Search

Momentum has been a consistently studied aspect of sports science for decades. Among the established literature, there has, at times, been a discrepancy between conclusions. However, if momentum is indeed an actual phenomenon, it would affect all aspects of sports, from player evaluation to pre-game prediction and betting. Therefore, using momentum-based features that quantify a team’s linear trend of play, we develop a data pipeline that uses a small sample of recent games to assess teams’ quality of play and measure the predictive power of momentum-based features versus the predictive power of more traditional frequency-based features across several leagues using several machine learning techniques. More precisely, we use our pipeline to determine the differences in the predictive power of momentum-based features and standard statistical features for the National Hockey League (NHL), National Basketball Association (NBA), and five major first-division European football leagues. Our findings show little evidence that momentum has superior predictive power in the NBA. Still, we found some instances of the effects of momentum on the NHL that produced better pre-game predictors, whereas we view a similar trend in European football/soccer. Our results indicate that momentum-based features combined with frequency-based features could improve pre-game prediction models and that, in the future, momentum should be studied more from a feature/performance indicator point-of-view and less from the view of the dependence of sequential outcomes, thus attempting to distance momentum from the binary view of winning and losing.

Traffic management in aviation, shipping, and rail transport shows similarities and dissimilarities in the work process. For example, they share the temporal aspect, but different levels of urgency in the control work set different requirements on monitoring, decisions, and actions. However, few studies have been presented that model and compare the different domains in terms of temporal decision-making. The Joint Control Framework (JCF) is an approach to analyse and temporally model operators’ control processes from a cognitive systems engineering perspective. In this study, we have used JCF to map, and compare, cognitive joints, such as perceptions, decisions, and actions, in temporally challenging control situations in air traffic control, maritime vessel traffic service, and train traffic management. Data was collected collaboratively with traffic operators, focusing on (1) identifying challenging traffic situations and (2) jointly modelling the temporal decision-making patterns of these situations using simplified JCF. Post-analysis was done by breaking down the results into different processes and comparing domains to ascertain how operators maintain control. An intermediate level of activity—between general monitoring and work with specific vehicles—was identified: processes-in-focus. A shared problem arises in the shift between general monitoring and the processes-in-focus. All processes-in-focus comprise cognitive joint cycles of perceptions, decisions, and actions. However, depending on the framing of processes-in-focus, the patterns of joints, such as temporal extension and complexity, differ. In the remainder of the article, implications for the interaction design, in particular the potential for human–AI/automation teaming with higher levels of automation and cognitive autonomy, are discussed., Funding: Swedish Transport Administration; [F-AUTO (TRV 2018/41347)]; [TRV 2020/138317]

In recent years, the metaverse—some form of immersive digital extension of the physical world—has received much attention. As tech companies present their bold visions, scientists and scholars have also turned to metaverse issues, from technological challenges via societal implications to profound philosophical questions. This article contributes to this growing literature by identifying the possibilities of two dystopian metaverse scenarios, namely one based on the experience machine and one based on demoktesis—two concepts from Nozick (Anarchy, State, and Utopia, Basic Books, 1974). These dystopian scenarios are introduced, and the potential for a metaverse to evolve into either of them is explained. The article is concluded with an argument for why the two dystopian scenarios are not strongly wedded to any particular theory of ethics or political philosophy, but constitute a more general contribution., Open access funding provided by RISE Research Institutes of Sweden. The author received no external funding for this work.

As more decisions are made by automated algorithmic systems, the transparency of these systems has come under scrutiny. While such transparency is typically seen as beneficial, there is a also a critical, Foucauldian account of it. From this perspective, worries have recently been articulated that algorithmic transparency can be used for manipulation, as part of a disciplinary power structure. Klenk (Philosophy & Technology 36, 79, 2023) recently argued that such manipulation should not be understood as exploitation of vulnerable victims, but rather as indifference to whether the information provided enhances decision-making by revealing reasons. This short commentary on Klenk uses Berlin’s (1958) two concepts of liberty to further illuminate the concept of transparency as manipulation, finding alignment between positive liberty and the critical account., Open access funding provided by RISE Research Institutes of Sweden. The author received noexternal funding for this work.

In many practical applications data used for training a machine learning model and the deployment data does not always preserve the same distribution. Transfer learning and, in particular, domain adaptation allows to overcome this issue, by adapting the source model to a new target data distribution and therefore generalizing the knowledge from source to target domain. In this work, we present a method that makes the adaptation process more transparent by providing two complementary explanation mechanisms. The first mechanism explains how the source and target distributions are aligned in the latent space of the domain adaptation model. The second mechanism provides descriptive explanations on how the decision boundary changes in the adapted model with respect to the source model. Along with a description of a method, we also provide initial results obtained on publicly available, real-life dataset. © The Author(s) 2024., Funding: The paper is funded from the XPM project funded by the National Science Centre, Poland under the CHIST-ERA programme (NCN UMO2020/02/Y/ST6/00070) and the Swedish Research Council under grant CHIST-ERA19-XAI-012 and by a grant from the Priority Research Area (DigiWorld) under the Strategic Programme Excellence Initiative at Jagiellonian University.

The digital product passport has been introduced as a policy instrument to enable traceability throughout the product life cycle and to support a circular economy. Anyhow, as a relatively new concept, there are a lot of uncertainties throughout the industrial landscape regarding the challenges and opportunities it brings. One such branch of industry is the furniture sector. In this paper, we present a DPP solution for furniture industry based on blockchain technology and inspired by a design thinking mindset. Through a prototype implementation, we highlight the key aspects such as data governance, stakeholder constellation, data interoperability, and data integrity that have significant research potential. Furthermore, we discuss these key concepts of DPP through the lens of eco-design principles in order to promote sustainability, improve energy efficiency, and protect the environment.

Assessing privacy risks and incorporating privacy measures from the onset requires a comprehensive understanding of potential impacts on data subjects. Privacy Impact Assessments (PIAs) offer a systematic methodology for such purposes, which are closely related to Data Protection Impact Assessments (DPIAs), particularly outlined in Article 35 of the General Data Protection Regulation (GDPR). The core of a PIA is a Privacy Risk Assessment (PRA). PRAs can be integrated as part of full-fledged PIAs or independently developed to support PIA processes. Although these methodologies have been identified as essential enablers of privacy by design, their effectiveness has been criticized because of the lack of evidence of their rigorous and systematic evaluation. Hence, we conducted a Systematic Literature Review (SLR) to identify published PIA and PRA methodologies and assess how and to what extent they have been scientifically validated or evaluated. We found that these methodologies are rarely evaluated for their performance in practice, and most of them have only been validated in limited studies. Most validation evidence is found with PRA methodologies. Of the evaluated methodologies, PIAs were the most evaluated, where case studies were the predominant evaluation method. These evaluated methodologies can be easily transferred to an industrial setting or used by practitioners, as they provide evidence of their use in practice. In addition, the findings in this study can be used to inform researchers of the current state-of-the-art, and practitioners can understand the benefits and current limitations of the methodologies and adopt evidence-based practices., Digital Health Innovation (DHINO) Project, DigitalWell Arena Project

For several decades now, safeguarding sensitive information from disclosure has been a key focus in computer science and information theory. Especially, in the past two decades, the subject of privacy has received significant attention due to the widespread collection and processing of data in various facets of society. A central question in this area is "What can be inferred about individuals from the data collected from them?" This doctoral thesis delves into a foundational and application-agnostic exploration of the theory of privacy. The overarching objective is to construct a comprehensive framework for evaluating and designing privacy-preserving data processing systems that adhere to three essential criteria: Explainability. The notion of information leakage (or privacy loss) employed in this framework should be operationally meaningful. That is, it should naturally emerge from the analysis of adversarial attack scenarios. Privacy guarantees within this framework should be comprehensible to stakeholders and the associated privacy parameters should be meaningful and interpretable. Robustness. The notion of information leakage employed should demonstrate resilience against a diverse array of potential adversaries, accommodating a broad range of attack scenarios while refraining from making restrictive assumptions about adversarial capabilities. Flexibility. The framework should offer value in a variety of contexts, catering to both highly privacy-sensitive applications and those with more relaxed privacy requirements. The notion of information leakage employed should also be applicable to various data types. The privacy notion proposed in this thesis that aligns with all the above criteria is called pointwise maximal leakage (PML). PML is a random variable that measures the amount of information leaking about a secret random variable X to a publicly available related random variable Y. We first develop PML for finite random variables by studying two seemingly d, Att skydda känslig information mot oavsiktligt avslöjande har varit ett viktigt forskningsmål inom datavetenskap och informationsteori under de senaste decennierna. I synnerhet under de senaste två decennierna har ämnet dataintegritet fått stor uppmärksamhet, inte minst på grund av den omfattande datainsamlingen som pågår i stora delar av samhället. En central fråga inom området är "Vilka slutsatser kan dras om individer från de data som samlas in från dem?" Denna avhandling fördjupar sig i teorin bakom dataintegritet från ett fundamentalt och tillämpningsoberoende perspektiv. Det övergripande målet är att skapa ett allsidigt ramverk för att designa och utvärdera dataintegritetsbevarande databehandlingssystem som följer tre essentiella kriterier: Förklarbarhet. Definitionen av informationsläckage (eller minskningen av dataintegritet) i detta ramverk bör ha en operationell betydelse, det vill säga att definitionen uppkommer naturligt från en analys av potentiella fientliga attacker. Dataintegritetsgarantier inom detta ramverk bör också vara förståeliga för intressenter, och motsvarande dataintegritetsparametrar bör vara meningsfulla och tolkningsbara. Robusthet. Definitionen av informationsläckage bör uppvisa motståndskraft mot en mångfald av potentiella fientliga attacker: definitionen bör vara tillämpbar på ett brett spektrum av fientliga attacker och undvika att göra restriktiva antaganden om den fientliga förmågan. Flexibilitet. Ramverket bör vara användbart i ett brett spektrum av tillämpningar; både i situationer där dataintegritet är av yttersta vikt, och där kraven inte är lika strikta. Definitionen av informationsläckage bör också vara applicerbart på olika datatyper. Definitionen av dataintegritet som presenteras i denna avhandling följer kriterierna ovan och kallas punktvist maximalt läckage (PML). PML är en stokastisk variabel som mäter mängden informationsläckage från en hemlig stokastisk variabel X till en relaterad, men publik, stokastisk variabel Y. V, QC 20240115

This study identifies the factors that guide the adoption of a public e-tourism system resulting in value co-creation in the UAE. Integrating and comparing factors drawn from the third version of the Technology Acceptance Model (TAM3), the Technology-Task-Fit (TTF) theory, and push-to-use, an Analytic Hierarch Process (AHP) model was implemented with data collected using a structured questionnaire from purposively selected UAE e-tourism experts (N = 15) and analyzed using Microsoft Excel. The findings revealed that usefulness, convenience of use, and push-to-use were the most critical aspects for achieving an efficient public e-tourism system that allows for value co-creation in that order of ranking. The findings also suggest that computer self-efficiency is the most critical factor in effectively establishing an e-tourism system followed by government push-to-use. In conclusion, the findings demonstrate that usefulness and ease-of-use backed by computer self-efficiency, result demonstrability, and output quality are vital for the efficient adoption of a public e-tourism system resulting in value co-creation in the UAE., Validerad;2024;Nivå 2;2024-04-09 (marisr);Funder: Slovenian Research Agency;Full text license: CC BY 4.0

Partial multi-label learning (PML) addresses problems where each instance is assigned a candidate label set and only a subset of these candidate labels is correct. The major challenge of PML is that the training procedure can be easily misguided by noisy labels. Current studies on PML have revealed two significant drawbacks. First, most of them do not sufficiently explore complex label correlations, which could improve the effectiveness of label disambiguation. Second, PML models heavily rely on prior assumptions, limiting their applicability to specific scenarios. In this work, we propose a novel method of PML based on the Encoder-Decoder Framework (PML-ED) to address the drawbacks. PML-ED initially achieves the distribution of label probability through a KNN label attention mechanism. It then adopts Conditional Layer Normalization (CLN) to extract the high-order label correlation and relaxes the prior assumption of label noise by introducing a universal Encoder-Decoder framework. This approach makes PML-ED not only more efficient compared to the state-of-the-art methods, but also capable of handling the data with large noisy labels across different domains. Experimental results on 28 benchmark datasets demonstrate that the proposed PML-ED model, when benchmarked against nine leading-edge PML algorithms, achieves the highest average ranking across five evaluation criteria.

User behavior is one of the biggest challenges to cybersecurity in modern organizations. Users are continuously targeted by attackers and required to have sufficient knowledge to spot and avoid such attacks. Different training methods are suggested and used in the industry to support users to behave securely. The challenge remains, and improved methods for end-user cybersecurity training are needed. This paper introduces and evaluates user perception of a method called Context-Based Micro-Training (CBMT). This approach suggests that training should be delivered in short sequences when the information is of direct relevance. The intention is to provide training directly related to the user’s current situation while also providing an awareness-increasing effect. This notion is tested in a survey-based evaluation involving 1,452 respondents from Sweden, Italy, and the UK, comparing the perception of CBMT against the experience of traditional approaches. The results emphasize that current methods are not effective enough and show that CBMT is perceived positively by respondents in all sample groups. The study further evaluated how demographic aspects impact the perception of CBMT and found that a diverse group of users can appreciate it., CC BY 4.0Published online: 09 Jun 2023CONTACT Joakim KävrestadThe work was supported by VINNOVA under the grant [2019-05021].

Artificial Intelligence (AI) and related technologies have a long history of being used in education for motivating learners and enhancing learning. However, there have also been critiques for a too uncritical and naïve implementation of AI in education (AIED) and the potential misuse of the technology. With the release of the virtual assistant ChatGPT from OpenAI, many educators and stakeholders were both amazed and horrified by the potential consequences for education. One field with a potential high impact of ChatGPT is programming education in Computer Science (CS), where creating assessments has long been a challenging task due to the vast amount of programming solutions and support on the Internet. This now appears to have been made even more challenging with ChatGPT’s ability to produce both complex and seemingly novel solutions to programming questions. With the support of data collected from interactions with ChatGPT during the spring semester of 2023, this position paper investigates the potential opportunities and threats of ChatGPT for programming education, guided by the question: What could the potential consequences of ChatGPT be for programming education? This paper applies a methodological approach inspired by analytic autoethnography to investigate, experiment, and understand a novel technology through personal experiences. Through this approach, the authors have documented their interactions with ChatGPT in field diaries during the spring semester of 2023. Topics for the questions have related to content and assessment in higher education programming courses. A total of 6 field diaries, with 82 interactions (1 interaction = 1 question + 1 answer) and additional reflection notes, have been collected and analysed with thematic analysis. The study finds that there are several opportunities and threats of ChatGPT for programming education. Some are to be expected, such as that the quality of the question and the details provided highly impact the qual

INTRODUCTION: Poor usability is a barrier to widespread adoption of electronic health records (EHR). Providing good usability is especially challenging in the health care context, as there is a wide variety of patient users. Usability benchmarking is an approach for improving usability by evaluating and comparing the strength and weaknesses of systems. The main purpose of this study is to benchmark usability of patient portals across countries. METHODS: A mixed-methods survey approach was applied to benchmark the national patient portals offering patient access to EHR in Estonia, Finland, Norway, and Sweden. These Nordic countries have similar public healthcare systems, and they are pioneers in offering patients access to EHR for several years. In a survey of 29,334 patients, both patients' quantitative ratings of usability and their qualitative descriptions of very positive and very negative peak experiences of portal use were collected. RESULTS: The usability scores ranged from good to fair level of usability. The narratives of very positive and very negative experiences included the benefits of the patient portals and experienced usability issues. The regression analysis of results showed that very positive and negative experiences of patient portal use explain 19-35% of the variation of usability scores in the four countries. The percentage of patients who reported very positive or very negative experiences in each country was unrelated to the usability scores across countries. CONCLUSIONS: The survey approach could be used to evaluate usability with a wide variety of users and it supported learning from comparison across the countries. The combination of quantitative and qualitative data provided an approximation of the level of the perceived usability, and identified usability issues to be improved and useful features that patients appreciate. Further work is needed to improve the comparability of the varied samples across countries., CC BY 4.0 DEEDCopyright © 2023 The Authors. Published by Elsevier B.V. All rights reserved.Corresponding author at: Department of Computer Science, Aalto University, Konemiehentie 2, Espoo FI-00076, Finland. E-mail address: sari.kujala@aalto.fi (S. Kujala).This work was supported by NordForsk through the funding to Nordic eHealth for Patients: Benchmarking and Developing for the Future, NORDeHEALTH, (Project #100477), the Swedish Research Council for Health, Working Life and Welfare (FORTE) through the funding to Beyond Implementation of eHealth (Project #2020-01229), the Strategic Research Council at the Academy of Finland (Project #352501 and #352503), and the Norwegian Centre for E-health Research.

Purpose - Online health question-and-answer (Q&A) forums have developed a new business model whereby listeners (peer patients) can pay to read health information derived from consultations between askers (focal patients) and answerers (physicians). However, research exploring the mechanism behind peer patients' purchase decisions and the specific nature of the information driving these decisions has remained limited. This study aims to develop a theoretical model for understanding how peer patients make such decisions based on limited information, i.e. the first question displayed in each focal patient-physician interaction record, considering argument quality (interrogative form and information details) and source credibility (patient experience of focal patients), including the contingent role of urgency. Design/methodology/approach - The model was tested by text mining 1,960 consultation records from a popular Chinese online health Q&A forum on the Yilu App. These records involved interactions between focal patients and physicians and were purchased by 447,718 peer patients seeking health-related information until this research. Findings - Patient experience embedded in focal patients' questions plays a significant role in inducing peer patients to purchase previous consultation records featuring exchanges between focal patients and physicians; in particular, increasingly detailed information is associated with a reduced probability of making a purchase. When focal patients demonstrate a high level of urgency, the effect of information details is weakened, while the interrogative form is strengthened. Originality/value - The originality of this study lies in its exploration of the monetization mechanism forming the trilateral relationship between askers (focal patients), answerers (physicians) and listeners (peer patients) in the business model “paying to view others' answers” in the online health Q&A forum and the moderating role of urgency in explai

Although the field of Artificial Intelligence in Education (AIEd) has a substantial history as a research domain, never before has the rapid evolution of AI applications in education sparked such prominent public discourse. Given the already rapidly growing AIEd literature base in higher education, now is the time to ensure that the field has a solid research and conceptual grounding. This review of reviews is the first comprehensive meta review to explore the scope and nature of AIEd in higher education (AIHEd) research, by synthesising secondary research (e.g., systematic reviews), indexed in the Web of Science, Scopus, ERIC, EBSCOHost, IEEE Xplore, ScienceDirect and ACM Digital Library, or captured through snowballing in OpenAlex, ResearchGate and Google Scholar. Reviews were included if they synthesised applications of AI solely in formal higher or continuing education, were published in English between 2018 and July 2023, were journal articles or full conference papers, and if they had a method section 66 publications were included for data extraction and synthesis in EPPI Reviewer, which were predominantly systematic reviews (66.7%), published by authors from North America (27.3%), conducted in teams (89.4%) in mostly domestic-only collaborations (71.2%). Findings show that these reviews mostly focused on AIHEd generally (47.0%) or Profiling and Prediction (28.8%) as thematic foci, however key findings indicated a predominance of the use of Adaptive Systems and Personalisation in higher education. Research gaps identified suggest a need for greater ethical, methodological, and contextual considerations within future research, alongside interdisciplinary approaches to AIHEd application. Suggestions are provided to guide future primary and secondary research.

För att kunna tillgängliggöra information till alla grupper i samhället behöver texter kun- na anpassas utifrån olika individers språkliga nivå. Idag används många gånger manu- ell textförenkling vilket är en väldigt resurskrävande metod. Det finns därför ett behov av automatiskt textförenklande system (Saggion & Hirst, 2017). Syftet med detta arbete var att utforska hur ett automatiskt textförenklande system kunde anpassas för personer med dyslexi och deras individuella språkliga nivå. Ett gemensamt drag för personer med dyslexi är att de har svårt för avkodningen av ord (”Dyslexiförbundet”, u. å). Förenkling av texter görs då genom att byta långa ord mot kortare (Saggion & Hirst, 2017). Den språk- liga nivån kan däremot variera och vissa individer hjälps därför även av att komplexa, lågt förekommande ord byts ut till enklare, mer frekventa ord (Rello, Baeza-Yates, Dempere- Marco & Saggion, 2013). Utifrån denna teori utvecklades ett lexikalt förenklande system som viktade ordens längd mot frekvens för att skapa personligt anpassade synonymut- byten. Systemet utvärderades genom att analysera hur viktningen påverkade valet av sy- nonym i ett utbyte. Resultatet visade att viktningsfunktionen kan generera synonymer av olika komplexitet beroende på hur parametern justerades. De olika synonymalternativen skulle i sin tur kunna användas för att skapa individuellt anpassade textförenklingar. Vi- dare forskning krävs däremot för att avgöra hur viktningsfunktionen ska balanseras samt vilka inställningar på viktningsparametrarna som lämpar sig för olika individer.

The growing demand for IT professionals to develop and maintain software is creating huge concerns for organizations looking to keep up and build resilience. The trends show that there is a large shortage compared to the demand. Manufacturing companies are no exception to this, and the adoption of Industry 4.0 technologies will most likely increase this need even further. However, developing software solutions often requires specialized knowledge, and organizations are looking far and wide for the competence. If it is present, it is often found in the organization’s centralized IT departments. Low-code development might offer a solution, by lowering the threshold for employees and individuals to develop, deploy and maintained applications with limited technical expertise. Low-code development is method that enables software applications to be built using graphical user interfaces in so called low-code development platforms. The platforms provide templates, drag-and-drop modules and integration with other systems to reduce the complexity of the application development process. These platforms could offer advantages for manufacturing firms, including quicker development that could aid in responding to new requirements. Other benefits include: cost savings and shorter development cycles, allowing for faster releases which can help enterprises stay more competitive and reduce risk. This thesis probes the factors to consider with low-code platforms, such as the role of citizen developers and the balance between accessibility and control in the context of manufacturing industries. It explores how low-code development could enhance efficiency and presents a case study from the engine assembly facility in Scania Södertälje. Best practices for building an organization around low-code developers are discussed. The insights from this study provide an understanding of the complexities and potential of low-code development, laying the groundwork for further exploration and imple

Flygresor.se, a leading flight comparison platform, uses machine learning to rankflights based on their likelihood of being clicked. The main goal of this project was toimprove this flight sorting to obtain a better user experience. The platform's existingmodel is based on a neural network approach and a limited set of features. The solution involved developing and comparing two machine learning models, Random Forest and XGBoost besides using a set of existing and newly created features. TheXGBoost model demonstrated superior performance by significantly improving theprediction of clicked flights by 4.18% while also achieving a remarkable increase inefficiency by being 125 times faster than the existing model., Flygresor.se, en ledande plattform för jämförelse av flygresor, använder maskininlärning för att ranka flygresor baserat på deras sannolikhet att bli klickade. Huvudmåletmed detta projekt var att förbättra denna flygsortering för att få en bättre användarupplevelse. Plattformens befintliga modell är baserad på ett neuralt nätverk och ettbegränsat antal funktioner. Lösningen innebar att utveckla och jämföra två maskininlärningsmodeller, Random Forest och XGBoost, förutom att använda en uppsättning befintliga och nyskapade funktioner. XGBoost-modellen visade bättre prestandagenom att förbättra predikteringen av de klickade flygresor med 4,18 % samtidigt somden uppnådde högre nivå av effektivitet genom att vara 125 gånger snabbare än denbefintliga modellen.

Den snabba tillväxten av uppkopplade enheter som samlar in och delar data leder till att användare kommer behöva läsa och godkänna alltmer sekretessmeddelanden för att kunna använda enheter och tjänster. Sekretessmeddelanden är de avtal som definierar villkoren för användning av en produkt eller tjänst och är avsedda att informera användare om insamlingen och användningen av deras data och personuppgifter innan valet att acceptera eller neka fattas. Sekretessmeddelanden tenderar vara långa samt innehålla mycket text med ett komplicerat språk. Texten i sekretessmeddelanden är ofta statisk utan användarinteraktion, vilket leder till att användare möts av en lång text att läsa utan någon form av stöd i form av designelement, exempelvis färg, ikoner, överblick eller interaktioner. Avsaknaden av designelement är en av orsakerna till att användare idag antingen förbiser eller accepterar sekretessmeddelanden utan att förstå innebörden, vilket får konsekvenser för användarens integritet. På grund av detta är det av stor vikt att användarna har en klar förståelse för hur deras integritet påverkas vid godkännande av ett sekretessmeddelande. Syftet med studien är att öka kunskapen inom området, och erbjuda och stötta designers i deras tillvägagångssätt kring hur sekretessmeddelanden kan designas för att främja användarens förståelse. Studien genomfördes genom en kvalitativ designstudie där ett befintligt ramverk skapat för design av sekretessmeddelanden kombinerades med identifierade designelement från litteraturstudien. Designelementen och ramverket användes vid framtagningen av fyra designförslag, vilka implementerades i en prototyp för att undersöka hur förståelse kan främjas i sekretessmeddelanden. Resultatet av studien är fyra designförslag som stöttar användaren att uppnå förståelse., The rapid growth of connected devices that collect and share data leads to users needing to read and approve an increasing number of privacy notices to use devices and services. Privacy notices are agreements that define the terms of use for a product or service and are intended to inform users about the collection and use of their data and personal information before deciding to accept or decline. Privacy notices tend to be lengthy and contain a lot of text with complicated language. The text in privacy notices is often static without user interaction, which leads to users being faced with a long text to read without any form of support in terms of design elements, such as color, icons, overview, or interactions. The lack of design elements is one of the reasons why users today either overlook or accept privacy notices without understanding the implications, which has consequences for the user's privacy. Because of this, it is of great importance that users have a clear understanding of how their privacy is affected by accepting a privacy notice. The purpose of the study is to increase knowledge in the field and offer and support designers in their approach to how privacy notices can be designed to promote user understanding. The study was conducted through a qualitative design study where an existing framework created for the design of privacy notices was combined with identified design elements from the literature study. The design elements and framework were used in the development of four design proposals, which were implemented in a prototype to investigate how understanding can be promoted in privacy notices. The result of the study is four design proposals that support users in achieving understanding.

The rarity and subtle radiographic features of atypical femoral fractures (AFF) make it difficult to distinguish radiologically from normal femoral fractures (NFF). Compared with NFF, AFF has subtle radiological features and is associated with the long-term use of bisphosphonates for the treatment of osteoporosis. Automatically classifying AFF and NFF not only helps improve the diagnosis rate of AFF but also helps patients receive timely treatment. In recent years, automatic classification technologies for AFF and NFF have continued to emerge, including but not limited to the use of convolutional neural networks (CNNs), vision transformers (ViTs), and multimodal deep learning prediction models. The above methods are all based on deep learning and require the use of centralized radiograph datasets. However, centralizing medical radiograph data involves issues such as patient privacy and data heterogeneity. Firstly, radiograph data is difficult to share among hospitals, and relevant laws or guidelines prohibit the dissemination of these data; Second, there were overall radiological differences among the different hospital datasets, and deep learning does not fully consider the fusion problem of these multi-source heterogeneous datasets. Based on federated learning, we implemented a distributed deep learning strategy to avoid the use of centralized datasets, thereby protecting the local radiograph datasets of medical institutions and patient privacy. To achieve this goal, we studied approximately 4000 images from 72 hospitals in Sweden, containing 206 AFF patients and 744 NFF patients. By dispersing the radiograph datasets of different hospitals across 3-5 nodes, we can simulate the real-world data distribution scenarios, train the local models of the nodes separately, and aggregate the global model, combined with percentile privacy protection, to further protect the security of the local datasets; in addition, we compare the performance of federated learning models us

Rapporten bygger på tre delstudier med olika datainsamlingstekniker: En bibliometristudie av svensk spelforskning baserad på data från Scopus respektive SwePub. En enkätstudie över svenska spelforskningsmiljöer. En intervjustudie om svenska spelforskarmiljöer. Bibliometristudiens ansats har varit bred. Materialet inkluderar därmed även poster som endast har en perifer koppling till spel, men returnerar i fallet SwePub färre än 3000 poster. Spelforskning bedrivs på många lärosäten i Sverige och publiceras inom många ämnesområden. Totalt 95 forskare har en publicering som omfattar 10 eller fler spelpublikationer1. Det finns starka lokala kluster av forskare som sampublicerar men sampubliceringen mellan svenska och nordiska lärosäten är inte så omfattande. Jämfört med Finland har Sverige färre publikationer med ett tydligt fokus på spelforskning. Enkätstudien visar att spelforskarmiljöerna oftast består av ett antal forskare med olika ämnesmässig bakgrund där de själva definierar sig som spelforskare snarare än att organisationen som sådan gör det. Miljöerna har sammantaget och högt räknat runt 30 aktiva doktorander men siffran är osäker. 94% av respondenterna anger att deras forskning ligger inom ”spel för andra områden än underhållning” och endast 6% anger ”spelteknikfokus”. 44% anger ”spelbranschen”. Det är tydligt att spel för just underhållning inte ligger fokus för svensk spelforskning. Detta är remarkabelt då den svenska datorspelsindustrin är en multimiljardindustri fokuserad på just spel som underhållning. Det finns alltså en mycket stor diskrepans mellan industri och svensk spelforskning. Intervjustudien ger en bild av att spelforskningen bedrivs inom splittrade miljöer, av forskare med olika ämnesmässig bakgrund. Det finns en tydlig upplevd brist på forskningsfinansiering inom adekvata områden och det finns brister i de finansieringsinstrument som finns idag. Detta påverkar såväl vilken typ av fo

This pilot study investigates how personalized content can be used in banking applications to encourage shaping good saving habits and increase overall financial literacy among young people. The preliminary result are recommendations including providing clear and personalized saving goals, incorporating educational content on investing and financial planning, and implementing features for tracking and categorizing expenses. Some trade-offs in usage that are presented are to minimize the use of personal data to what is necessary for personalization purposes, treat the data on a group level or increase transparency of data usage. The pilot study concludes that personalization is likely to be beneficial for both banks and their customers, given personal data is handled carefully and used in a sound financial consumer protection framework.

Confronted with the aggression against Ukraine in 2022, Europe faces one of the most important humanitarian challenges - the migration of war refugees from Ukraine, most of them women with children and the elderly. Both international institutions such as the European Union and the United Nations, but also national governments and, above all, local governments, which are the main providers of services and resources for refugees, are taking a number of measures to meet the needs. The extraordinary nature and extensive humanitarian needs pose exceptional challenges for both governments and Non-Governmental Organizations (NGOs) as well as civil society. The European countries adopted distinct reception procedures to accommodate war refugees in their territories. The purpose of this paper is to examine the challenges of war refugees from Ukraine and gain an understanding of how they vary across selected European countries. Using a text analytics approach such as BERTopic topic modeling, we analyzed text messages published on Telegram channels from February 2022 to September 2023, revealing 12 challenges facing Ukrainian migrants. Furthermore, our study delves into these challenges distribution across 6 major European countries with significant migrant populations, providing insights into regional differences. Additionally, temporal changes in 8 narrative themes in discussions of Ukrainian migration, extracted from official government websites, were examined. Together, this research contributes (1) to demonstrating how analytics-driven methodology can potentially be used to extract in-depth knowledge from textual data freely available on social media; and (2) to a deeper understanding of the various issues affecting the adaptation of Ukrainian migrants in European countries. The study also provides recommendations to improve programs and policies to better support the successful integration of Ukrainian migrants in host countries.

Given the importance of identifying and monitoring news stories within the continuous flow of news articles, this paper presents PromptStream, a novel method for unsupervised news story discovery. In order to identify coherent and comprehensive stories across the stream, it is crucial to create article representations that incorporate as much topic-related information from the articles as possible. PromptStream constructs these article embeddings using cloze-style prompting. These representations continually adjust to the evolving context of the news stream through self-supervised learning, employing a contrastive loss and a memory of the most confident article-story assignments from the most recent days. Extensive experiments with real news datasets highlight the notable performance of our model, establishing a new state of the art. Additionally, we delve into selected news stories to reveal how the model’s structuring of the article stream aligns with story progression., Also part of series: LREC proceedings, ISBN: 2522-2686

The rise of Artificial Intelligence has put data and data quality at the core of digitalisation. At the same time there seems to be a need to better understand what is meant by data quality and how to ensure it is at hand. Our first attempt to start the discussion was made at the Data Space Symposium in Darmstadt in March 2024. In terms of contribution this report is an increment to the slideshow by giving further examples of how data quality is defined and put into use at the same time as highlighting the contextual properties of data quality and the need for human judgement. We do this from a policy perspective, i.e. grounding our analysis in regulations and standards. Our analysis starts with the legal reasonings on data quality found in the AI Act and the European Health Data Space regulation. Our ambition is not to be exhaustive, there are more EU regulations and directives to consider in relation to data quality than the ones we cover here – such as the directive on Copyright in the Digital Single Market that introduces the concept of text and data mining; the Data and Data Governance Acts that enable standardised formats for making data interoperable across services; and the Digital Service and Market Acts that define responsibilities in terms of making data and information available. Among others. The same goes for standards on data quality which span across specific domains and disciplines. Coming back to our ambition, we believe it is important to raise the question to what extent data quality can be automatically assessed, as this is an ambition floated at various events and foras within the data community. While we think this can be achieved for specific and narrow contexts, we argue that data quality is a topic that still requires judgement and the competence to make assessments on a case-by-case basis.

This report focus on the intersection ofForeign Information Manipulation andInterference and Large Language Models.The aim is to give a non-technicalcomprehensive understanding of howweaknesses in the language models canbe used for creating malicious content tobe used in FIMI.

The future role of AI in healthcare is largely seen as positive, inevitable, necessary, and driven by the private sector. These dominant narratives obscure decision-making about these future applications, with little room for political scrutiny, or public participation. Public, and private resources, are being funnelled to support various AI health technologies. By prioritising some and not others, we are being set on a path that will greatly impact our future health, lives, and societies. My research sheds light on these processes and provides tools to expand the scope for inclusive public participation in the debate on future role of AI in healthcare.

Modern society is increasingly dependent on digital services, making their dependability a top priority. But while there is a consensus that cybersecurity is important, there is no corresponding agreement on the true extent of the problem, the most effective countermeasures, or the proper division of labor and responsibilities. This makes cybersecurity policy very difficult. This article addresses this issue based on observations and experiences from a period of guest research at the Swedish Financial Supervisory Authority (Finansinspektionen), which made it possible to study how cybersecurity policy is developed and implemented in the Swedish financial sector. Observations include policy implementation challenges related to squaring different roles and perspectives mandated by different laws, and to collaboration between independent government authorities, but also policy development challenges: How can the full range of perspectives and tools be included in cybersecurity policy development? As Sweden now revises its cybersecurity policy, this is a key issue., Den aktuella studien är finansierad av Stiftelsen för Strategisk Forskning (avtalsnummer SM22-0057)QC 20240807

In this thesis we studied synaptic plasticity and neuronal computation in single striatal projection neurons (SPNs), which have a major role in goal-directed learning. Goal-directed or reward learning means to learn, based on sensory information from the body and the environment, to select actions out of all the behavioral repertoire that lead to obtaining a goal or reward (such as food or water). In mammals, all the behavioral motor repertoire is under constant, tonic inhibition, and the direct-pathway SPNs (dSPNs) select (disinhibit) goal-obtaining actions. The learning process is guided by the neuromodulator dopamine which signals the positive or negative out-come of an action. The synapses from cortical neurons on to the dSPNs, called corticostriatal synapses, are responsive to dopamine signals, and can strengthen and weaken based on the (positive or negative) action outcome. This promotes or discourages future actions in the same or similar sensory context. Within a collaborative computational modeling effort, we studied the biochemical circuitry in the corticostriatal synapses with multiscale modeling and simulations. This circuitry in the corticostriatal synapses responds to neuromodulatory signals and controls the expression of synaptic plasticity. Multiscale modeling and simulations enable studying a system at multiple temporal and spatial scales, and integrating the results across the different scales. Based on molecular dynamics simulations of the enzyme which transduces extracellular neuromodulatory signals into an intracellular second messenger molecule, and Brownian dynamics simulations of regulator molecules binding to the enzyme, we constructed a kinetic model ofthe enzyme-based signal transduction network. The kinetic model showed that two co-occuring neuromodulatory signals, a dopamine peak and an acetylcholine pause, are required to produce the second messenger and thus enable strengthening of corticostriatal synapses onto dSPNs, and that only the, I denna avhandling studerades synaptisk plasticitet samt förmågan till lokalberäkning i de striatala projektionsneuronens (SPNs) dendriter. Detta har direkt relevans för förståelsen av s.k. målstyrd inlärning. Målstyrd inlärning och belöningsinlärning innebär att man, baserat på sensorisk informationfrån kroppen och omgivningen, lär sig selektera hur man skall agera/handla så att man uppnår ett mål eller erhåller en belöning (t.ex. mat eller vatten). Hos alla däggdjur är motoriska/exekutiva centra i hjärnstammen och thalamus under konstant tonisk inhibition via basala ganglierna, men när manaktiverar SPNs i den s.k. direkta vägen genom basala ganglierna (dSPN) så disinhiberas de motoriska centra som behövs för att initiera specifika mål-styrda beteenden. Inlärningsprocessen för detta guidas av bl.a. dopamin, en neuromodulator som signalerar huruvida ett beteende ger positivt eller negativt resultat. Synapserna från kortikala projektionsneuron till dSPNs, s.k. kortikostriatala synapser, påverkas av dopamin och de kan förstärkas eller försvagas baserat på om dopaminsignalen signalerar ett positivt eller negativt utfall. Detta antingen främjar eller motverkar att man väljer samma beteende/handlingar under en liknande situation i framtiden. Via samarbete med andra beräkningsbiologigrupper kunde vi studera de biokemiska signaleringsnätverken i de kortikostriatala synapserna med hjälp av multiskal modeller och simuleringar. Synapsens intracellulära biokemiska signaleringsnätverk kontrollerar synaptisk plasticitet och påverkas av neuro-modulering. Genom att använda multiskalsimuleringar kunde vi studera systemet över multipla temporala och spatiala skalor, och integrera resultaten över de olika skalorna. Baserat på molekylärdynamiska simuleringar av det enzym som överför de extracellulära neuromodulatoriska signalerna till en intracellulär s.k. 2nd messenger molekyl, och Brownianska dynamiksimuleringar av de regulatoriska molekyler som binder till enzymet, kunde vi konstru, QC 20240904

In the paper Sparks of Artificial General Intelligence, the authors show how OpenAI’s GPT-4 is able do well in variety of tasks that be represented with text and claim it to have “a more general intelligence than previous AI models.” One of the tasks they explore is symbolic music generation. In this paper we critically analyze their results and extend the discourse around the capabilities of LLMs for music by exploring additional musical tasks and LLMs. Furthermore, we will investigate the viability of smaller models when used in conjunction with Retrieval Augmented Generation, as well as finetuning on programmatically written prompts using Quantized Low Rank Adapters. Finally, we discuss some critical aspects of LLMs as a tool for music generation., QC 20240906

This paper introduces an AI-based approach to detect human-made objects and changes in these on land parcels. To this end, we used binary image classification performed by a convolutional neural network. Binary classification requires the selection of a decision boundary, and we provided a deterministic method for this selection. Furthermore, we varied different parameters to improve the performance of our approach, leading to a true positive rate of 91.3% and a true negative rate of 63.0%. A specific application of our work supports the administration of agricultural land parcels eligible for subsidiaries. As a result of our findings, authorities could reduce the effort involved in the detection of human made changes by approximately 50%.

Visual Learning Analytics (VLA) uses analytics to monitor and assess educational data by combining visual and automated analysis to provide educational explanations. Such tools could aid teachers in primary and secondary schools in making pedagogical decisions, however, the evidence of their effectiveness and benefits is still limited. With this scoping review, we provide a comprehensive overview of related research on proposed VLA methods, as well as identifying any gaps in the literature that could assist in describing new and helpful directions to the field. This review searched all relevant articles in five accessible databases — Scopus, Web of Science, ERIC, ACM, and IEEE Xplore — using 40 keywords. These studies were mapped, categorized, and summarized based on their objectives, the collected data, the intervention approaches employed, and the results obtained. The results determined what affordances the VLA tools allowed, what kind of visualizations were used to inform teachers and students, and, more importantly, positive evidence of educational interventions. We conclude that there are moderate-to-clear learning improvements within the limit of the studies’ interventions to support the use of VLA tools. More systematic research is needed to determine whether any learning gains are translated into long-term improvements.

Schools and educational institutions collect large amounts of data about students and their learning, including text, grades, quizzes, timestamps, and other activities. However, in primary and secondary education, this data is often dispersed across different digital platforms, lacking standardized methods for collection, processing, analysis, and presentation. These issues hinder teachers and students from making informed decisions or strategic and effective use of data. This presents a significant obstacle to progress in education and the effective development of Educational Technology (EdTech) products. Visual Learning Analytics (VLA) tools, also known as Learning Analytics Dashboards (LADs), are designed to visualize student data to support pedagogical decision-making. Despite their potential, the effectiveness of these tools remains limited. Addressing these challenges requires both technical solutions and thoughtful design considerations, as explored in Papers 1 through 5 of this thesis. Paper 1 examines the design aspects of VLA tools by evaluating higher education data and various visualization and Machine Learning (ML) techniques. Paper 2 provides broader insights into the VLA landscape through a systematic review, mapping key concepts and research gaps in VLA and emphasizing the potential of VLA tools to enhance pedagogical decisions and learning outcomes. Meanwhile, Paper 3 delves into a technical solution (data pipeline and data standard) considering a secure Swedish warehouse, SUNET. This includes a data standard for integrating educational data into SUNET, along with customized scripts to reformat, merge, and hash multiple student datasets. Papers 4 and 5 focus on design aspects, with Paper 4 discussing the proposed Human-Centered Design (HCD) approach involving teachers in co-designing a simple VLA tool. Paper 5 introduces a scenario-based framework for Multiple Learning Analytics Dashboards (MLADs) development, stressing user engagement for tailored

Continuous long-term monitoring of motor health is crucial for the early detection of abnormalities such as bearing faults (up to 51% of motor failures are attributed to bearing faults). Despite numerous methodologies proposed for bearing fault detection, most of them require normal (healthy) and abnormal (faulty) data for training. Even with the recent deep learning (DL) methodologies trained on the labeled data from the same machine, the classification accuracy significantly deteriorates when one or few conditions are altered, e.g., a different speed or load, or for different fault types/severities with sensors placed in different locations. Furthermore, their performance suffers significantly or may entirely fail when they are tested on another machine with entirely different healthy and faulty signal patterns. To address this need, in this pilot study, we propose a zero -shot bearing fault detection method that can detect any fault on a new (target) machine regardless of the working conditions, sensor parameters, or fault characteristics. To accomplish this objective, a 1D Operational Generative Adversarial Network (Op-GAN) first characterizes the transition between normal and fault vibration signals of (a) source machine(s) under various conditions, sensor parameters, and fault types. Then for a target machine, the potential faulty signals can be generated, and over its actual healthy and synthesized faulty signals, a compact, and lightweight 1D Self-ONN fault detector can then be trained to detect the real faulty condition in real time whenever it occurs. To validate the proposed approach, a new benchmark dataset is created using two different motors working under different conditions and sensor locations. Experimental results demonstrate that this novel approach can accurately detect any bearing fault achieving an average recall rate of around 89% and 95% on two target machines regardless of its type, severity, and location.

The need for handling the deep uncertainty surrounding the future climate has led to various novel and robust approaches for decision-making under deep uncertainty (DMDU) when adapting to climate change. Here, an online and self-explanatory web-based tool was developed and tested with civil servants from five municipalities in Sweden challenged by rising sea levels. The municipalities used the tool by themselves and were then interviewed about the usability of the tool, the perceived urgency of climate change adaptation, and the possibilities for municipalities for handling the flexible solutions that are at the core of DMDU. Results were that the civil servants reported that the urgency of climate change adaptation has increased, that the tool changed their perspectives compared to traditional planning methods, and that changes in laws and regulations to accommodate decisions about flexible solutions were recommended.

This research focuses on the analysing spontaneous infant movements using Hidden Markov Model (HMM) based analysis. While the usual research in this field addresses the lack of automated, scalable, and objective methods for assessing these movements to predict neurodevelopmental disorders such as cerebral palsy (CP), our research aims to analyse these movements from a more holistic perspective. The aim of our research is to utilize VI-HMMs to find out more about infant movement as well as find the optimal parameters for such models to enable futher research. This problem is significant and suitable for a Bachelor’s thesis due to its potential impact on understanding of early motor and neural development. Diagnosis and intervention for CP, a condition affecting 2 to 3 per 1000 live births is one of the clearest applications for this research. Early diagnosis and intervention can significantly improve outcomes for affected children, making this a critical area of study . The challenge lies in the complexity of accurately modeling and analyzing spontaneous movements, a task that has not been fully addressed due to the limitations in computational techniques and resources. The methodology involved using Variational Bayesian Inference Hidden Markov Models (VI-HMM) to analyze spatio-temporal sequences of infant movements. This approach was chosen for its ability to handle the multivariate and probabilistic nature of the data, providing a robust framework for identifying underlying patterns in the movements. The study utilized a dataset of infant movement sequences, extracting and normalizing various representations of the data for analysis. Evaluation was performed using cross-validation and information criteria to select the best models and states . We did not succeed in fulfilling our first goal, that being finding out more about infant movements themselves. We did however manage to fulfill our second goal, that being finding the optimal parameters. The main implication, Denna forskning fokuserar på att analysera spontana spädbarnsrörelser med hjälp av Hidden Markov Model (HMM) baserad analys. Medan den vanliga forskningen inom detta område tar upp bristen på automatiserade, skalbara och objektiva metoder för att bedöma dessa rörelser för att förutsäga neuroutvecklingsstörningar som cerebral pares (CP), syftar vår forskning till att analysera dessa rörelser ur ett mer holistiskt perspektiv. Syftet med vår forskning är att använda VI-HMM för att ta reda på mer om spädbarns rörelser samt hitta de optimala parametrarna för sådana modeller för att möjliggöra vidare forskning. Detta problem är betydande och lämpligt för en kandidatuppsats på grund av dess potentiella inverkan på förståelsen av tidig motorisk och neural utveckling. Diagnos och intervention för CP, ett tillstånd som påverkar 2 till 3 per 1000 levande födda, är en av de tydligaste tillämpningarna för denna forskning. Tidig diagnos och intervention kan avsevärt förbättra resultaten för drabbade barn, vilket gör detta till ett kritiskt studieområde. Utmaningen ligger i komplexiteten i att noggrant modellera och analysera spontana rörelser, en uppgift som inte har åtgärdats helt på grund av begränsningarna i beräkningstekniker och resurser. Metodiken involverade att använda Variational Bayesian Inference Hidden Markov Models (VI-HMM) för att analysera spatio-temporala sekvenser av spädbarnsrörelser. Detta tillvägagångssätt valdes för dess förmåga att hantera datas multivariata och probabilistiska karaktär, vilket ger ett robust ramverk för att identifiera underliggande mönster i rörelserna. Studien använde en datauppsättning av spädbarns rörelsesekvenser, extraherade och normaliserade olika representationer av data för analys. Utvärderingen utfördes med hjälp av korsvalidering och informationskriterier för att välja de bästa modellerna och tillstånden. Vi lyckades inte uppfylla vårt första mål, att själva ta reda på mer om spädbarnsrörelser. Vi lyckades dock uppfylla vårt andr

Environmental factors can subtly shape our motor skills from infancy, where infants transition from spontaneous to coordinated movements. This early development, largely unaffected by varied surroundings, offers a unique insight into the natural evolution of motor abilities. This thesis utilises unsupervised machine learning (UML) to analyse spontaneous infant movements through preprocessed video-recorded data. Principal Component Analysis (PCA) and autoencoders were applied to break down the complex data into more manageable parts. Following this, k-means clustering was used to try to identify distinct patterns. A Random Forest Classifier and silhouette score was used to analyse the clusters, along with a heatmap to analyse the dimensionality reduction. Although the clustering did not yield well-defined groups, it uncovered significant insights into the variability and complexity of infant movements. These observations align with General Movements Assessment (GMA) principles, which also recognise movement variability as a marker of neurological health. The absence of distinct movement patterns might be attributed to the infants’ limited exposure to environmental influences., Faktorer i miljön kan subtilt forma våra motoriska färdigheter redan från spädbarnsåldern, där spädbarn övergår från spontana till koordinerade rörelser. Denna tidiga utveckling, som till stor del inte har hunnit påverkas av olika omgivningar, erbjuder en unik inblick i de naturliga utvecklingsstegen för motoriska förmågor. Denna avhandling använder oövervakad maskininlärning (UML) för att analysera spontana spädbarnsrörelser genom förbehandlad videodata. Principalkomponentsanalys (PCA) och autoencoders tillämpades för att förenkla de komplexa datamängderna till mer hanterbara delar. Efter detta användes k-means klusteranalys för att försöka identifiera distinkta mönster. En Random Forest-Klassificerare och silhuettvärde användes för att analysera klustren, tillsammans med en värmekarta för att undersöka dimensionsreduktionerna. Även om klustringen inte resulterade i väldefinierade grupper, avslöjade den betydande insikter i variationen och komplexiteten i spädbarnsrörelserna. Dessa observationer framhäver principerna för General Movements Assessment (GMA), som också betonar rörelsevariabilitet som en markör för neurologisk hälsa. Spädbarnens brist på distinkta rörelsemönster kan möjligen förklaras av deras begränsade exponering av miljöfaktorer.

This report investigates the use of machine learning to predict customer churn in the insurance industry from a broker’s perspective. Given the high effort required to acquire new customers compared to retaining existing ones, minimizing customer churn is crucial. This study, conducted with Söderberg & Partners, aims to understand and predict customer churn using existing data. Machine learning models, including logistic regression, random forest, and gradient boosting, were developed and evaluated for their accuracy, precision, and recall. The results indicate that advanced models like XGBoost outperform simpler models in predicting churn with high accuracy and precision. This research also examines the economic impact of these models, demonstrating potential cost savings and improved customer retention strategies. The study highlights the ethical considerations in handling sensitive data, ensuring compliance with GDPR. This work provides valuable insights for the insurance sector, contributing to both academic knowledge and practical improvements in broker-client relationships and operational efficiency., Denna rapport undersöker användningen av maskininlärning för att förutsäga kundavhopp i försäkringsbranschen från en förmedlares perspektiv. Med tanke på den höga insats som krävs för att skaffa nya kunder jämfört med att behålla befintliga, är det avgörande att minimera kundavhopp. Denna studie, genomförd tillsammans med Söderberg & Partners, syftar till att förstå och förutsäga kundavhopp med hjälp av befintlig data. Maskininlärningsmodeller, inklusive logistisk regression, random forest och gradient boosting, utvecklades och utvärderades för deras noggrannhet, precision och återkallelse. Resultaten visar att avancerade modeller som XGBoost överträffar enklare modeller i att förutsäga kundavhopp med hög noggrannhet och precision. Forskningen undersöker också den ekonomiska påverkan av dessa modeller och visar på potentiella kostnadsbesparingar och förbättrade strategier för kundbevarande. Studien belyser de etiska övervägandena vid hantering av känslig data och säkerställer efterlevnad av GDPR. Detta arbete ger värdefulla insikter för försäkringssektorn och bidrar till både akademisk kunskap och praktiska förbättringar i förmedlar-kundrelationer och operativ effektivitet.

To understand how the brain processes information, we must first determine how information is represented in the neural activity. One way to do this is by exposing subjects to stimuli and see how their brains react. The goal of this paper is to either support or undermine previous research and their findings. According to this previous research, the noise in the brain of subjects should decrease when they are exposed to visual stimulation. The computations were made the same way with the same statistical model of distribution as the previous research, with the main difference being the dataset which contained a different species of animal. This dataset was derived from the Allen Brain Atlas. A key finding in this paper is that the neurons in the brain are already active at the stimulus onset, and that there is great trial-to-trial variability in both pre-stimulus and stimulus evoked activity. While the majority of findings were equivalent to the ones in the compared research, we did also examine some outliers where the noise in the brain stayed flat or even increased. Although, these can in great length be explained by trial-to-trial variations and other relevant research theories which suggest this kind of behaviour., För att förstå hur hjärnan bearbetar information måste vi först fastställa hur information representeras i den neurala aktiviteten. Ett sätt att göra detta är genom att utsätta objekt för stimulans och se hur deras hjärnor reagerar. I denna uppsats använder vi oss av tidigare forskning för att antingen stödja eller motbevisa deras resultat. Enligt den tidigare forskningen borde bruset i objektens hjärnor minska när de utsätts för visuell stimulans. Beräkningarna gjordes på samma sätt med samma statistiska distributionsmodell som den tidigare forskningen. Den enda skillnaden var att vårt dataset inkluderade en annan djurart. Datasetet hämtades från Allen Brain Atlas. En huvudupptäckt i det här arbetet är att hjärnans neuroner redan var aktiva vid stimulansappliceringen. Arbetet visar även att det finns en stor variabilitet från test till test vid aktivitet skapad av förstimulans såväl som stimulans. Fastän majoriteten av våra upptäckter liknar de som finns i den tidigare forskningen, noterade vi även vissa avvikande värden där bruset antingen förblev detsamma eller till och med ökade. Dock kan dessa generellt förklaras av variationer mellan försöken och vissa andra forskningsteorier vilka tyder på den här typen av beteende.

Traceability is important in any production facility. Alleima, a producer of stainless steel and alloys will also benefit from this by keeping track of its products. In this thesis, a billet punch mark system was developed to identify billets that arrive at an in-transit location and accurately detects and predicts the billet punch mark numbers. The system also includes a simple GUI (Graphical User Interface) which can be used at the facility. The punch marks will be saved along with the corresponding time of prediction, helping Alleima trace and correct any issues that might arise during production. Many steps were developed as part of this study to identify the punch marks: 1) template matching was used to find and crop the billet and billet holder, 2) SAD (Sum of Absolute Differences) was used to find a frame for each billet, 3) Canny filtering along with the Hough transform was applied to crop out the billet, 4) a YOLO (You Only Look Once) object detection model was trained to locate the punch mark region of interest (ROI), 5) the ROI was rotated using two rotation schemes, and 6) the numbers were read by OCR (Optical Character Recognition) software PaddleOCR. For the final evaluation, seven videos were processed. Out of a total of 292 billets, all were found, and the punch mark numbers (or lack thereof) were accurately detected in 92.5% of the cases., Spårbarhet är en viktig aspekt inom tillverkningsindustrin. Alleima, ett företag som tillverkar rostfritt stål och legeringar, kan också nyttja detta genom att lättare hålla ordning på sina produkter. I detta examensarbete skapades ett system för att finna stålbalkar som anländer till en specifik plats under produktion, för att sedan kunna läsa av deras stansmärken. Systemet har även ett enklare grafiskt användargränssnitt för att programmet lättare ska kunna användas av personal på anläggningen. Siffrorna som läses av kommer att sparas tillsammans med tiden de lästes, detta kan hjälpa Alleima spåra och korrigera eventuella fel som kan uppståt under produktionen. Flertalet steg utvecklades för att läsa dessa stansmärket: 1) mallmatchning för att hitta och segmentera järnbalken och hållaren, 2) SAD (Summan av absoluta skillnader) användes för att hitta en bildruta för varje järnbalk, 3) Canny filtrering användes tillsammans med Hough transformen for att segmentera kortsidan på järnbalken, 4) en YOLO objektidentifiering modell tränades för att hitta stansmärkesområdet, 5) dessa områden roterades korrekt med hjälp av två olika rotationssystem, 6) siffrorna läses av en OCR modell (Optisk igenkänningsmodell) för att identifiera siffrorna. För att evaluera systemet testades detta system på olika MP4 filer. Sju olika videor bearbetades. Alla 272 järnbalkar hittades, dess stansmärkesnummer (eller avsaknaden av sådana) kunde korrekt identifieras i totalt 92.5 % av fallen.

Latency is a crucial aspect of object recognition systems, being fast enough to recognize objects allows the use of these algorithms for various industrial, security, or robotics applications that require real-time detection. A typical example of a real-time application of object detection is self-driving cars, whereby algorithms need to recognize obstacles as quickly as possible, to avoid collisions. This thesis investigates how it is possible to reduce the latency during the recall of neural network-based object detection systems deployed on an embedded system. It focuses on measuring the latency of the various steps required to run an object detection algorithm on a security camera, exploring an approach to reduce the latency, and evaluating at which cost in terms of accuracy these improvements in speed comes. The main attempt to reduce the latency proposed in this thesis is to skip the Image Processing Pipeline (IPP), which is the process that transforms the raw data obtained from the sensor into a more common RGB image and trains the object detection system to work directly on the raw data obtained from the sensors. This approach allows skipping entirely the processing pipeline and spares that computational time, letting the network instead learn how to deal directly with raw images. The final results are promising, however, the reduction in latency greatly depends on the hardware used to capture the images. A security RGB camera produced by Axis will be used as an example for the latency evaluation, and the Coco [1] dataset to evaluate the accuracy., Latens är en avgörande aspekt för objektigenkänning system, att vara tillräckligt snabb för att känna igen objekt gör det möjligt att använda dessa algoritmer för olika industri-, säkerhets- eller robottillämpningar som kräver realtidsdetektering. Ett typiskt exempel på en realtidsapplikation för objektdetektering är självkörande bilar, där algoritmerna måste känna igen hinder så snabbt som möjligt för att undvika kollisioner. Denna avhandling undersöker hur det är möjligt att minska latensen under återkallandet av neurala nätverksbaserade objektdetekteringssystem som används på ett inbyggt system. Den fokuserar på att mäta latensen för de olika steg som krävs för att köra en algoritm för objektdetektering på en säkerhetskamera, utforska en metod för att minska latensen och utvärdera till vilken kostnad i form av noggrannhet dessa förbättringar i hastighet kommer. Det huvudsakliga försöket att minska latensen som föreslås i denna avhandling är att hoppa över Image Processing Pipelin (IPP), vilket är den process som omvandlar rådata från sensorn till en mer vanlig RGB-bild och tränar objektdetekteringssystemet att arbeta direkt med rådata från sensorerna. Detta tillvägagångssätt gör det möjligt att helt hoppa över bearbetningspipelinen och spara den beräkningstiden, så att nätverket istället kan lära sig att direkt hantera råa bilder. De slutliga resultaten är lovande, men minskningen av latenstiden beror i hög grad på vilken hårdvara som används för att ta bilderna. En RGB-säkerhetskamera från Axis kommer att användas som exempel för att utvärdera latensen, och Coco-datasetet[1] för att utvärdera noggrannheten.

Coreference resolution is an area within natural language processing, namely the task of determining which expressions in a text refer to the same entity. It is a useful resource for other natural languages processing activities, such as text summarization. Although the latest methods within the field of coreference resolution utilize deep learning, earlier rule-based methods can still be relevant in certain cases. One of the rule-based coreference resolution models developed for German is CoRefGer-rule. It implements the Stanford sieve algorithm originally developed for other languages, such as English. The authors reported that CoRefGer-rule was useful especially for out-of-domain texts where no gold standard information was available, arguing the relevancy of rule-based models. They also gave improvement suggestions for their model. One of the suggestion was to add a morphological analyzer in order to make use of grammatical information about words. Another suggestion was to implement a rule that considers the semantic aspect of words, i.e. the aspect related to the meaning of words. In this project, we developed our own rule-based coreference resolution model based on CoRefGer-rule, where these two suggestions were examined. One of our conclusions was that the use of grammatical information obtained from the morphological analyzer was a fundamental and important part of our system, enabling advanced resolution of pronouns. On the other hand, the semantic information had no significant impact on the performance of our model., Koreferenslösning är ett språkteknologiskt område som handlar om att avgöra vilka uttryck i en text som refererar till samma fenomen i världen. Koreferenslösning kan i sin tur användas för andra språkteknologiska syften, t.ex. automatisk textsammanfattning. De senaste metoderna för koreferenslösning använder sig av djupinlärning, men även äldre regelbaserade metoder kan vara lämpliga för användning i vissa fall. En av de regelbaserade modellerna som utvecklats för tyska är CoRefGer-rule. Den bygger på den s.k. Stanford sieve-algoritmen som ursprungligen utvecklades för ett antal andra språk, bl.a. engelska. Utvecklarna av CoRefGer-rule menade att deras modell var användbar i synnerhet för texter utanför domänen som saknade guldstandardsdata. De gav även förslag på hur deras modell skulle kunna förbättras. Ett av förslagen var att bygga in morfologisk analys för att kunna utnyttja grammatisk information om ord. Ett annat förslag var att lägga till en regel som tar hänsyn till ordsemantik, alltså ordens betydelse. Det här projektet gick ut på att utveckla ett system inspirerat av CoRefGer-rule och inkludera de funktioner som nämndes i förbättringsförslagen ovan. En av våra slutsatser var att den grammatiska information som extraherades från den morfologiska analysen var en fundamental och viktig del av vårt system. Den möjliggjorde bl.a. avancerad koreferenslösning för pronomen. Samtidigt visade sig den semantiska informationen inte ha någon särskild påverkan på resultatet.

This research constitutes an empirical exploration into the application of generative artificial intelligence (AI) and machine learning methodologies for predictive analysis and data visualization, with a concentrated focus on how Generative AI can generate accurate and meaningful results based on specific questionnaires such as on the prediction of housing prices in California. The comparative study evaluates the efficacy of OpenAI for Microsoft Azure with Lida and Mulesoft OpenAI platforms in executing these analytical tasks. The results of the predictive analysis revealed a clear association between geographic location and housing prices, with properties designated as ’<1H OCEAN’ showing significantly higher median values as opposed to those located in ’INLAND’ areas. While Azure OpenAI (Lida) demonstrated expertise in generating inferential statistics using specific datasets, Mulesoft OpenAI assisted in conducting statistical analysis and identifying potential predictors. Mulesoft’s integration with OpenAI saw noticeable growth with generative AI functionality, despite the absence of data visualization capabilities. A specific instance highlighting the impact of dataset size on results was observed in the examination of the average cost of living in California across different platforms. With 480 rows of datasets, Azure OpenAI (Lida) indicated an average living cost of $185,665.97, while the same dataset size for MuleSoft OpenAI yielded a significantly different result of $246,057. This discrepancy suggests that MuleSoft OpenAI may have limitations in handling larger datasets, potentially impacting the accuracy of its results. In contrast, Azure OpenAI’s ability to handle larger datasets, as evidenced by the result of $251,000 when using the full CHPD, indicates the importance of sufficient data for more accurate predictions. These findings emphasize the need for careful consideration of platform capabilities and dataset sizes when conducting data analysis ta, Denna forskning utgör en empirisk utforskning av tillämpningen av generativ artificiell intelligens (AI) och maskininlärningsmetoder för prediktiv analys och datavisualisering, med ett fokuserat intresse för hur Generativ AI kan generera precisa och meningsfulla resultat baserade på specifika frågeformulär, såsom förutsägelser om bostadspriser i Kalifornien. Den jämförande studien utvärderar effektiviteten hos OpenAI för Microsoft Azure med Lida och Mulesoft OpenAI-plattformarna när det gäller att utföra dessa analytiska uppgifter. Resultaten av den prediktiva analysen avslöjade en tydlig koppling mellan geografisk plats och bostadspriser, där fastigheter betecknade som ’<1H OCEAN’ visade betydligt högre medianvärden jämfört med de som är belägna i ’INLAND’-områden. Medan Azure OpenAI (Lida) visade expertis i att generera inferentiell statistik med specifika dataset, erbjöd Mulesoft OpenAI hjälp med att genomföra statistisk analys och identifiera potentiella prediktorer. Mulesofts integration med OpenAI såg märkbar tillväxt med generativ AIfunktionalitet, trots frånvaron av datavisualiseringsförmåga. Ett specifikt exempel som belyser datasetets storlekens påverkan på resultat observerades vid undersökningen av den genomsnittliga levnadskostnaden i Kalifornien över olika plattformar. Med 480 rader av datasetet angav Azure OpenAI (Lida) en genomsnittlig levnadskostnad på $185665,97, medan samma datasetstorlek för MuleSoft OpenAI gav ett betydligt annorlunda resultat på $246 057. Denna diskrepans antyder att MuleSoft OpenAI kan ha begränsningar i att hantera större dataset, vilket potentiellt påverkar noggrannheten i dess resultat. Å andra sidan indikerar Azure OpenAI:s förmåga att hantera större dataset, som bevisas av resultatet på $251000 när hela Kaliforniens bostadsdata används, vikten av tillräcklig data för mer exakta förutsägelser. Dessa resultat betonar behovet av noggrant övervägande av plattformars förmågor och datasetstorlekar vid utförandet av dataanaly

This study investigates the key factors influencing Swedish inflation and employs a Long Short-Term Memory (LSTM) based machine learning model to identify the most contributing factors through prediction. Using data from Swedish National Bureau of Statistics, the research focuses on aspects such as state finances, demand, supply and expectations. The LSTM based model was chosen for its ability to handle sequence data and long-term dependencies, making it suitable for economic forecasting. Monthly data from 2000 to 2023 were collected and normalized to ensure consistency. The LSTM based model was trained and validated using different configurations. The most accurate predictions were achieved using seven factors: previous inflation, money supply, GDP, national debt, exchange rates (Euro and Dollar), and capacity utilization, yielding a Mean Absolute Error (MAE) of 0.0491. Remarkably, even with just three factors (previous inflation, money supply, and GDP), the model maintained a high predictive accuracy with an MAE of 0.0515. The findings provide valuable insights for policymakers, particularly the Riksbank and the Swedish Government. Identifying the primary drivers of inflation supports more informed decision-making and proactive measures to stabilize the economy. Furthermore, this research aligns with global development goals by promoting sustainable economic growth and innovation through better financial planning and resource allocation., Denna studie undersöker bidragande faktorer som påverkar svensk inflation och använder en Long Short-Term Memory (LSTM)-baserad maskininlärningsmodell för att identifiera dessa faktorer genom prediktion. Data från Statistiska centralbyrån används, med fokus på aspekter som statsfinanser, utbud, efterfrågan förväntningar. Den LSTM-baserade modellen valdes för dess förmåga att hantera sekvensdata och långsiktiga beroenden, vilket gör den lämplig för ekonomiska prognoser. Månadsdata från 2000 till 2023 samlades in och normaliserades. Den LSTM-baserade modellen tränades och validerades med olika konfigurationer. De mest exakta prediktionerna uppnåddes med sju faktorer: tidigare inflation, penningmängd, BNP, statsskuld, valutakurser (Euro och Dollar) och kapacitetsutnyttjande, vilket gav ett medelabsolutfel (MAE) på 0,0594. Även med bara tre faktorer (tidigare inflation, penningmängd och BNP) upprätthöll modellen hög prediktiv noggrannhet med ett MAE på 0,0491. Resultaten ger värdefulla insikter för beslutsfattare, särskilt Sveriges Riksbank och Sveriges regering. Att identifiera de primära drivkrafterna bakom inflationen stödjer mer informerade beslut och proaktiva åtgärder för att stabilisera ekonomin. Vidare främjar denna forskning hållbar ekonomisk tillväxt och innovation genom bättre finansiell planering och resursallokering, i linje med globala utvecklingsmål.

The amount of people with dementia in Sweden is increasing, as are the associated costs. Thus, technological innovations, digitalization, and new work practices are becoming more crucial to accommodate this development. There is an increasing demand for a more time-effective dementia diagnosis process, and one area for improvement is within speech assessments. This thesis has investigated whether it is possible to develop an interface for multidimensional clinical language analysis of dementia. The result was an automated tool that evaluates transcribed speech from several different dimensions. These include a machine learning prediction, a syntactic analysis in the form of a word dependency plot, a semantic analysis, and a series of different linguistic features. The visualization was created with the Python package Gradio and divided into two tabs: one with the numerical values and one with graphs to facilitate comparisons to baselines from the data. Furthermore, a market analysis has been conducted to investigate the potential of such an application in the Swedish medical market and where it could be implemented. This was conducted through interviews with three different market participants. From the interviews, it was found that there is potential for such an application for speech-language pathologists, who conduct more thorough speech assessments. The interviews revealed that the analysis of speech is still heavily manual-intensive and presents an opportunity for time optimization. It is in this aspect that the market viability of the tool lies., Antalet personer i Sverige med demens ökar, liksom kostnaderna inom demensvården. Tekniska innovationer, digitalisering och nya arbetssätt blir därför allt mer avgörande för att utrusta sjukvården för denna förändring. Det finns ett ökat behov av en mer tidseffektiv demensutredning och ett förbättringsområde är inom talbedömningar. Detta examensarbete har undersökt om det är möjligt att utveckla ett gränssnitt för multidimensionell klinisk språkanalys av demens. Resultatet blev ett automatiserat verktyg som utvärderar transkriberat tal från flera olika dimensioner. Detta inkluderar en maskininlärningsmodell, en syntaktisk analys i form av ett språkträd, en semantisk analys och en utvärdering av olika språkliga egenskaper. Visualiseringen var skapad med hjälp av Python-paketet Gradio och delades upp i två flikar: en med numeriska värden och en med grafer för att underlätta jämförelser med referensvärden från datan. Vidare har en marknadsanalys genomförts för att undersöka potentialen för en sådan applikation inom den svenska vården och var verktyget skulle kunna implementeras. Denna analys har genomförts genom intervjuer med tre olika marknadsaktörer. Man fann att det finns potential för ett sådant verktyg för logopeder, som utför mer noggranna talbedömningar. Intervjuerna avslöjade att analysen av tal fortfarande är en manuellt intensiv process som lämnar utrymme för tidsoptimering och att det är i denna aspekt som verktygets marknadspotential finns.

When studying gene expression in cells with single cell mRNA sequencing (scRNA-seq), the gene expression data is of a very high dimension. The amount of data that is being generated in a single scRNA-seq experiment to study gene expression is increasing fast, therefore the amount of computational resources needed for analyses is also increasing. Furthermore, it is difficult to get a meaningful overview of high dimensional data. To solve these issues, one often applies dimensionality reduction through Principal Component Analysis (PCA) that compresses the data into a low dimensional form. However, PCA relies on linear assumptions, and biological data is often not linear. In this thesis, an AutoEncoder was trained to see if it can perform better than PCA, as AutoEncoders does not rely on assumptions of linearity. A systematic search for the best performing architecture of the AutoEncoder was performed, and this model was then compared against PCA for dimensionality reduction from 500 dimensions to 100, 75, 50, 25, 10 and 5 dimensions. The AutoEncoder performed better than PCA for the higher dimensions (for which it was trained on during the architectural search), but worse in the lower dimensions. As such, the training of AutoEncoders has great potential to improve the low-dimensional fidelity of high-dimensional data, though it might be challenging to generalize a model that works for a large range of lower dimensional compression. Therefore, while AutoEncoders show promise, their application requires careful consideration and optimization., När man studerar genuttryck i celler med single cell mRNA-sekvensering (scRNA-seq) är genuttrycksdata av mycket hög dimensionalitet. Mängden data som genereras i varje scRNA-seq-experiment för att studera genuttryck ökar snabbt, vilket medför ett ökande behov av beräkningsresurser för analys. Vidare är det dessutom svårt att få en meningsfull överblick av högdimensionell data. För att lösa dessa problem tillämpar man ofta dimensionsreduktion genom Principal Component Analysis (PCA) som komprimerar data till en lågdimensionell form. PCA bygger dock på linjära antaganden, vilket ofta inte håller när man arbetar med biologisk data. I detta projekt tränades en AutoEncoder för att undersöka om den kan prestera bättre än PCA, eftersom AutoEncoders inte bygger på antaganden om linjäritet. En systematisk sökning efter den bäst presterande arkitekturen av AutoEncodern genomfördes, och denna modell jämfördes sedan med PCA för dimensionsreduktion från 500 dimensioner till 100, 75, 50, 25, 10 och 5 dimensioner. AutoEncodern presterade bättre än PCA för de högre dimensionerna (som den tränades på under arkitektursökningen), men sämre i de lägre dimensionerna. Därmed har träning av AutoEncoders stor potential att förbättra den lågdimensionella representationen av högdimensionell data, men det kan vara svårt att träna en generell AutoEncoder som fungerar för ett större omfång av lågdimensionella kompressioner. AutoEncoders visar lovande resultat, men deras tillämpning kräver noggrann övervägning och optimering.

This study investigates the preprocessing and hyperparameter tuning of support vector machine (SVM) classifiers applied to high-dimensional, small sample-sized, and incomplete (HSI) datasets, with a specific focus on classifying Duchenne Muscular Dystrophy (DMD) disease progression using proteomic data. Our research aimed to establish a methodological framework to evaluate the reliability and performance of SVM classifiers in this context. The data preprocessing steps included feature selection using Recursive Feature Elimination with Cross-Validation (RFECV), data imputation via scikit-learn’s IterativeImputer with RandomForestRegressor as the estimator, and data normalization using both Z-score and min-max normalization. To optimize the SVM models, an exhaustive grid search for hyperparameter tuning was performed with 5-fold cross-validation. The evaluation metrics included ROC-AUC, F 1 , precision, recall, and accuracy. The SVM models did not achieve clinically significant predictions despite this systematic methodological framework. The primary challenges identified were class imbalance, model instability, and overfitting. The high-dimensional and incomplete nature of the proteomic data, combined with the experimental nature of the IterativeImputer, likely introduced biases that impacted the overall model performance. Moving forward, it is essential to expand the dataset and adapt the methodology by incorporating domain-specific insights. Validating identified biomarkers by fundamental biomarker research and expanding the dataset with more diverse cohorts are promising ways to refine, as well as confirm the clinical relevance of, selected features and classifier models. It is possible that alternative machine learning models and advanced imputation strategies could address the limitations of the current methodology in the context of this specific HSI dataset, although the current dataset quality might still render modeling an impossible challenge. Overall, this, Denna studie undersöker förbehandling och hyperparameterjustering av stödvektormaskin (SVM) klassificerare tillämpade på högdimensionella, små och ofullständiga (HSI) dataset, med ett specifikt fokus på klassificering av sjukdomsprogression för Duchennes muskeldystrofi (DMD) med hjälp av proteomisk data. Vår forskning syftade till att etablera ett metodologiskt ramverk för att utvärdera tillförlitligheten och prestandan hos SVM-klassificerare i denna kontext. Databehandlingsstegen inkluderade variabelselektion med Recursive Feature Elimination with Cross-Validation (RFECV), dataimputation via scikit-learn’s IterativeImputer med RandomForestRegressor som estimator och datanormalisering med både Z-score och min-max normaliseringstekniker. En omfattande grid search för hyperparameterjustering utfördes med 5-faldig korsvalidering för att optimera SVM-modellerna. Utvärderingsmåtten inkluderade ROC-AUC, F1, precision, recall och accuracy. Trots en noggrann implementering av detta ramverk uppnådde SVM-modellerna inte kliniskt signifikanta förutsägelser. De huvudsakliga utmaningarna som identifierades var klassobalans, modellinstabilitet och överanpassning. Den högdimensionella och ofullständiga naturen av proteomikdatan, kombinerat med IterativeImputer’s experimentella status, introducerade sannolikt bias som påverkade den övergripande modellprestandan. Framåt är det väsentligt att utöka datan och fortsatt anpassa metodologin genom att införliva domänspecifika insikter. Validering av identifierade biomarkörer genom grundläggande biomarkörforskning samt utökning av datan med mer varierade kohorter är lovande vägar framåt för att förfina, samt bekräfta den kliniska relevansen hos, de utvalda proteinerna och den resulterande klassificeringsmodellen. Det är möjligt att alternativa maskininlärningsmodeller och avancerade imputationstekniker skulle kunna adressera begränsningarna i den nuvarande metodologin i relation till detta specifika HSI-dataset. Samtidigt är det också full

In recent years, rapid advancements in machine learning and artificial intelligence have spurred the development of specialized hardware designed to optimize performance and efficiency. Among these, Neural Processing Units (NPUs) have emerged as a promising technology for accelerating neural network tasks. This study compares the efficiency of an NPU with a Central Processing Unit (CPU) and Graphical Processing Unit (GPU) in executing an object detection model, YOLOv5. The comparison focuses on inference time, power, and energy consumption. Pre-trained YOLOv5 models of various sizes, precisions, and formats are used to fairly showcase the capabilities of each hardware platform. The YOLOv5 models used in this study come in nano, small, medium, and large sizes. The models on the CPU and GPU are available in single and low precisions, in both PyTorch and ONNX formats. The models on the NPU are in half and low precisions, each in RKNN format. The results highlight the NPU’s superiority in power and energy efficiency, though it is slower than the PyTorch models on the GPU in terms of inference time. PyTorch models on the GPU are the fastest and the second most energy-efficient, while being the most power-intensive. The CPU serves as a middle-ground option in terms of inference time and power consumption, while being the most energy-intensive., Under de senaste åren har snabba framsteg inom maskininlärning och artificiell intelligens drivit utvecklingen av specialiserad hårdvara som är utformad för att optimera prestanda och effektivitet. Bland dessa har neurala processorenheter (NPUs) framstått som en lovande teknologi för att påskynda neurala nätverksuppgifter. Denna studie jämför effektiviteten hos en NPU med en centralprocessor (CPU) och en grafisk processor (GPU) vid exekvering av en objektigenkänningsmodell, YOLOv5. Jämförelsen fokuserar på inferenstid, effekt och energiförbrukning. Förtränade YOLOv5-modeller i olika storlekar, precisioner och format används för att rättvist visa kapaciteten hos varje hårdvaruplattform. YOLOv5-modellerna som används i denna studie finns i storlekarna nano, liten, medium och stor. Modellerna på CPU och GPU finns i enkel och låg precision, i både PyTorch- och ONNX-format. Modellerna på NPU är i halv och låg precision, alla i RKNN-format. Resultaten framhäver NPU överlägsenhet i effekt och energieffektivitet, även om den är långsammare än PyTorch-modellerna på GPU när det gäller inferenstid. PyTorch-modellerna på GPU är de snabbaste och näst mest energieffektiva, men mest effektkrävande. CPU fungerar som ett mellanalternativ när det gäller inferenstid och effektförbrukning, men är mest energiintensiv.

Designing software is not only about creating clever algorithms to solve complex problems better and faster. In today's digital landscape, web developers must also make conscious decisions when producing systems in order not to compromise their users in case of an attack to steal data or harm the website. This is where Secure by Design comes in, a design principle aimed at fostering a more proactive approach to development by creating robust frameworks with rigid foundations that conform to the application unique rules, requirements, and constraints in which the program will be used. In this thesis, WordPress will be examined under the lens of how it integrates certain Secure by Design concepts, giving the reader an insight into how large-scale software is designed to be secure. The investigation was done by examining the open-source code base of WordPress and previous vulnerabilities that had compromised the application's security. The results indicated that WordPress understands its domain well and how corresponding files with these domain entities should interact. It seems like the vulnerabilities were all caused by overlooking the cleaning of user input during development. This led to the conclusion that WordPress should consider adopting a more modular approach like the one used by Secure by Design, given the available resources., Att designa mjukvara handlar inte bara om att skapa smarta algoritmer för att lösa komplexa problem bättre och snabbare. I dagens digitala landskap måste webbutvecklare också göra medvetna beslut när de producerar system för att inte äventyra sina användare vid en attack som syftar till att stjäla data eller skada webbplatsen. Det är här konceptet \textit{Secure by Design} kommer in, en designprincip med syftet att få utvecklare röra sig mer till en proaktiv ställning i deras utveckling vad gäller säkerhet genom att designa robusta strukturer med rigida underliggande strukturer som följer den domän och dess regler där programmet ska användas. WordPress kommer i denna avhandling att undersökas ur perspektivet hur det integrerar vissa Secure by Design-koncept, vilket ger läsaren insikt i hur storskalig mjukvara designas för att vara säker. Undersökningen gjordes genom en granskning av den öppna källkoden till applikationen, samt tidigare sårbarheter som äventyrat applikationens säkerhet. Resultaten från detta visade att WordPress har en god förståelse för sin domän och hur motsvarande filer med dessa domänentiteter ska interagera. Sårbarheterna uppstod alla från försummelse i städning av användarinmatning under utveckling, vilket ledde till slutsatsen att det bästa sättet framåt för WordPress möjligtvis är att anta det mer modulära tillvägagångssätt som Secure by Design erbjuder, givet att resurserna finns där.