Search

Machine Learning (ML) and Automated Machine Learning (Auto-ML) had attracted more attention and interest in industrial applications in recent years. Machine learning pipeline includes repetitive tasks such as data pre-processing, feature engineering, model selection and hyper-parameter optimization. Building a machine learning model requires extensive time for development, stress-testing, and multiple experiments. Besides, building a model with a small search space of pipeline steps with multiple algorithms and hyper-parameters takes hours. Hence, Auto-ML has been widely adapted to save time and efforts on such tasks. Nevertheless, there are many limitations in the existing Auto-ML frameworks. Most of the existing Auto-ML pipeline frameworks focus only on a part of the ML pipeline which does not always result in the optimum overall pipeline for the given dataset. Moreover, several Auto-ML pipeline frameworks ignore integrating meta-learning, thus they recommend a well-performing pipeline for a single task and not a global and generic optimal pipeline. Thus, for new tasks they must search for a new pipeline. Furthermore, although there are Auto-ML frameworks that consider the complete pipeline but imbalanced datasets did not receive much attention from the proposed Auto-ML frameworks. To tackle these challenges, we propose a new efficient Auto-ML framework that is equipped with a meta learning mechanism for automated algorithm selection which also handles imbalanced datasets efficiently. This paper also demonstrates how the proposed framework outperforms the-state-of-the-art frameworks.

Internet of Things (IoT) is a network which connects different communication devices with the internet to attain quick, robust and real-time information transfer and communication, achieving intelligent management. IoT is still in its infancy so it faces numerous challenges varying from data management to security concerns. Sensors generate enormous quantities of data that need to be handled efficiently to have successful deployment of IoT applications. Concerning data management, a great challenge that faces the IoT environment is the detection of contextual anomalies. Contextual anomaly detection is a sophisticated task because the context has to be taken into consideration in the anomaly detection process rather than checking only the deviation of the data value as in point anomaly detection. As a result, in this paper, a novel clustering based algorithm is proposed to detect contextual anomalies in Internet of Things. Attributes were separated into two different categories, namely contextual attributes and behavioral attributes. K-Means clustering technique was applied on the contextual and behavioral attributes separately, then the intersection between the contextual and behavioral clusters was used to detect the contextual anomalies. Moreover, the algorithm was applied on a real room occupation dataset of size around 20,000 records and the experiments showed promising results.

People’s opinions are considered as the most powerful source of market research. Popularly, Social Media has become a tool that is used in an easy way including huge number of users who can share their opinions about products or services and their thoughts about current problems of the society and express their views on political and religious issues. The knowledge extracted from social media contains sentiment data – that is not included in corporate database – that can be used to improve the marketing campaigns to retain customers and meet their needs in a better way. The integration and merging between both social media data and corporate data can lead to better insights that would not have been possible to gain without such integration. In this paper, we will use Twitter as a social media source platform to do a feature based level sentiment analysis using tweets including opinions about a specific product. . The research discussed three different ways to extract (feature/opinion) pairs from each text including: Normal Tokenization, N-gram Modeling Extraction, and Noun Chunking Extraction. The extracted opinion phrase related to each extracted feature is being classified using sentiment classification algorithm. A decision is taken about the best between the three ways according to the resulted measurements. SCDJF had been evaluated using multiple techniques. The best results occurred from Noun Chunking Extraction with accuracy 77%. Summarization of the results will show how this can be used to enhance decision making process of the organization. Summarization of the results will show how this can be used to enhance decision making process of the organization.

Internet of Things (IoT) is a network which connects several devices with the internet to have quick and real-time information transmission, attaining intelligent management. IoT is still in its beginning so it faces a lot of challenges. One huge challenge related to data management is the detection of collective anomalies. A collective anomaly is a set of points which differ as a whole from the rest of the data points. However, the individual points are not considered anomalies. Detection of collective anomalies is of huge importance as collective anomalies usually denote the occurrence of events or certain incidents that should be detected as soon as possible before they cause damage to the surrounding environment. Detection of collective anomalies is considered a complicated process as it involves more than one point rather than only one point as in point anomaly detection process. As a result, in this paper, we propose a novel approach of collective anomaly detection using a variable sized window in IoT. Two techniques were used to detect the collective anomalies, namely: K-Means clustering and isolation forests. The proposed approach was applied on a machine temperature dataset of size around 22,000 and the experiments showed promising results.

In this paper we study the problem of finding semantic relationships between folksonomy tags. We investigate different methods used to embed tags in the vector space and find similarities between them using word embedding vectors. We also present two new methods for embedding tags in the vector space utilizing labeled Latent Dirichlet Allocation (LDA) and Wikipedia category links. Related tags are grouped into communities using an overlapping community detection technique. In order to evaluate tag embedding methods, we use three different evaluation metrics, two of them do not require a ground truth dataset and the third is based on a manually created dataset of ground truth communities. Our results show that representing folksonomy tags using bag of words and embedding this representation in the vector space yields the best results compared to embedding co-occurring tags only or embedding tags along with textual content of tagged documents. We also compare between using word embedding, Latent Semantic Indexing (LSI), and LDA to find similarities between bag of words representations of tags. We show that word embedding outperforms LSI in one representation, while LDA is hard to beat.

In the present generation, on-Line social networks (OSNs) have become increasingly popular, people’s social lives have become more associated with these sites. They use on-Line social networks (OSNs) to keep in touch with each others, share news, organize events, and even run their own e-business. The rapid growth of OSNs and the massive amount of personal data of its subscribers have attracted attackers, and imposters to steal personal data, share false news, and spread malicious activities. On the other hand, researchers have started to investigate efficient techniques to detect abnormal activities and fake accounts relying on accounts features, and classification algorithms. However, some of the account’s exploited features have negative contribution in the final results or have no impact, also using standalone classification algorithms does not always achieve satisfactory results. In this paper, a new algorithm, SVM-NN, is proposed to provide efficient detection for fake Twitter accounts and bots, feature selection and dimension reduction techniques were applied. Machine learning classification algorithms were used to decide the target accounts identity real or fake, those algorithms were support vector machine (SVM), neural Network (NN), and our newly developed algorithm, SVM-NN. The proposed algorithm (SVM-NN) uses less number of features, while still being able to correctly classify about 98% of the accounts of our training dataset.

Online Communities for Question Answering (CQA) such as Quora and Stack Overflow face the challenge of providing sufficient answers for the questions asked by users. The exponential growing rate of the unanswered questions compromises the effectiveness of the CQA frameworks as knowledge sharing platforms. The main reason for this issue is the inefficient routing of the questions to the potential answerers, who are the field experts and interested users. This paper proposes the deep-learning-based technique QR-DSSM to increase the accuracy of the question routing process. This technique uses deep semantic similarity model (DSSM) to extract semantic similarity features using deep neural networks and use the features to rank users' profiles. QR-DSSM maps the asked questions and the profiles of the users into a latent semantic space where the ability to answer is measured using the cosine similarity between the questions and the profiles of the users. QR-DSSM experiments outperformed the baseline models such as LDA, SVM, and Rank-SVM techniques and achieved an MRR score of 0.1737.

The use of social media platforms such as social networks, weblogs and community question answering websites increased largely in the last few years. This increase in usage contributed to the vast explosion in available online content. Some of these platforms use social tagging, tags manually inserted by content authors, as a way to facilitate content description and discovery. In this paper, our goal is to automatically group semantically related tags in order to organize the large amount of tags contributed by various users. Our approach is based on using a topic model to discover topics of documents, and then grouping top tags related to documents assigned to each topic. We perform a set of experiments using different number of topics to provide different levels of details for the generated tag groups. The dataset used in our experiments is extracted from Stack Overflow, a community question answering website used by programming professionals. Tag groups generated by our technique are presented and evaluated.

Online Communities for Question Answering (CQA) such as Quora and Stack Overflow face the challenge of providing high quality answers to the questions asked by their users. Although CQA frameworks receive new questions in a linear rate, the rate of the unanswered questions increases in an exponential way. This variation eventually compromise effectiveness of the CQA frameworks as knowledge sharing platforms. The main cause for this challenge is the improper routing of questions to the potential answerers, field experts or interested users. The proposed technique QR-DSSM uses deep semantic similarity model (DSSM) to extract semantic similarity features using deep neural networks. The extracted semantic features are used to rank the profiles of the answerers by their relevance the routed question. QR-DSSM maps the asked questions and the profiles of the users into a latent semantic space where the relevance is measured using cosine similarity between the two; questions and users’ profiles. QR-DSSM achieved MRR score of 0.1737. QR-DSSM outperformed the baseline models such as query likelihood language model (QLLM), Latent Dirichlet Allocation (LDA), SVM classification technique and RankingSVM learning to rank technique.

A considerable amount of attributes in real datasets are not numerical, but rather textual and categorical. We investigate the problem of identifying outliers in categorical and textual datasets. We propose a clustering-based semi-supervised outlier detection method which basically represents normal and unlabeled data points as a bipartite graph. We leverage the existing free of parameters clustering techniques to cluster the resulting graph. The bipartite graph is clustered with a specific end goal to distinguish unlabeled data points as either outliers or normal data points. The proposed method is evaluated using multiple categorical and textual datasets against one-class support vector machines classifier and FRaC approach for semi-supervised outlier detection where the proposed method has shown a comparable performance.

Opinions are considered a powerful source of market research. Social networks are the most popular mean for people to state their opinions about products and services. Companies invest in analysing their customers' opinions from multiple existing social media platforms. The knowledge extracted from social media contains sentiment data that is not included in corporate database. This extracted data can be used to improve marketing campaigns to better retain customers and meet their needs. The integration between both social media data and corporate data can lead to better insights that would not have been possible to gain without such integration. This paper proposes a framework to integrate sentiment data, extracted from customers' opinions with corporate data of an organization. The process of integrating opinions streams into corporate database is proposed, then a multidimensional data warehouse is built over the integrated database to perform advanced analytical tasks and answer queries that would not have been possible without the integration. The new framework has been experimented with a case study of integrating opinions about digital devices and corporate data to illustrate the usage of the integrated model.

The rapid increase in data on social media creates a need for mining such data to get valuable insights. The data type can be unstructured with large volumes. Sentiment analysis addresses such need by detecting opinions or emotions on the social media text. Sentiment analysis can be performed in various domains such as social, medical and industrial applications. This paper presents a survey about sentiment analysis addressing the different concepts in this area, problems and its solutions, available APIs, tools used and presenting a list of open challenges in this area.

Managing temporal data gets increased demand for many significant areas such as scientific and financial applications. This paper proposes a new index (TOIX), which is designed in order to provide more efficient evaluation of temporal queries on XML documents. The index lies on mapping the twigs into temporal objects and then using these objects instead for answering the query. An improvement of the naive algorithm using a B-tree as well. Finally, a set of conducted experiments were performed and showed that our proposed algorithm outperforms the state of the art indexing algorithms in certain cases.

Social media contains a lot of useful information that express different aspects of the society using unstructured and temporal data. The main challenge with social media is the huge amount of the data streams to be processed. Such huge data requires huge processing capabilities, including the power, time and cost.This research presents a framework to extract, analyze and visualize the demographics of any country using the micro blogger data streams such as twitter. This increases the ability to search for topics with hash tags per city for a specific country as well as applying data mining techniques. The proposed framework uses big data tools including Hadoop and Hive to analyze the huge amount of textual unstructured data that is extracted from twitter then visualize it as needed.

Different models have been proposed recently for representing temporal data, tracking historical information and retrieving temporal queries results efficiently. We consider the problem of indexing temporal XML documents. In particular, we propose an indexing scheme that uses a summary structure and a matrix that captures the structural relationships as well as time intervals inside a temporal XML document. We introduce an algorithm to efficiently process all types of temporal queries with any depth using our newly proposed index. We show that our proposed index out-performs the state of the art indices in terms of both query processing time and support for different temporal query types.

Inspired by the great success of information retrieval (IR) style keyword search on the web, keyword search on XML has emerged recently. Existing methods cannot resolve challenges addressed by using keyword search in Temporal XML documents. We propose a way to evaluate temporal keyword search queries over Temporal XML documents. Moreover, we propose a new ranking method based on the time-aware IR ranking methods to rank temporal keyword search queries results. Extensive experiments have been conducted to show the effectiveness of our approach.

We introduce a new technique for fast computation of structural join "pattern trees" in XML. Using a small amount of pre-computed path information, typically small enough to fit easily in main memory, we are able to render structural join computation almost independent of data set size. Our technique is amenable to bit-mapped processing, leading to further speed-up. In this paper, we present our technique and experimentally evaluate its performance.

We propose a new algorithm, BPI-TWIG , to evaluate XML twig queries. The algorithm uses a set of novel twig indices to reduce the number of comparisons needed for the twig evaluation and transform the join operation to an intersection operation between the contributing twig paths inside the query. In this paper, we present our technique and experimentally evaluate its performance.