Search

Machine learning research, particularly in genomics, is often based on wide shaped datasets, i.e. datasets having a large number of features, but a small number of samples. Such configurations raise the possibility of chance influence (the increase of measured accuracy due to chance correlations) on the learning process and the evaluation results. Prior research underlined the problem of generalization of models obtained based on such data. In this paper, we investigate the influence of chance on prediction and show its significant effects on wide shaped datasets. First, we empirically demonstrate how significant the influence of chance in such datasets is by showing that prediction models trained on thousands of randomly generated datasets can achieve high accuracy. This is the case even when using cross-validation. We then provide a formal analysis of chance influence and design formal chance influence estimators based on the dataset parameters, namely its sample size, the number of features, the number of classes and the class distribution. Finally, we provide an in-depth discussion of the formal analysis including applications of the findings and recommendations on chance influence mitigation.

Machine learning may enhance prediction of outcomes after coronary artery bypass grafting (CABG). We sought to develop and validate a dynamic machine learning model to predict CABG outcomes at clinically relevant pre- and postoperative time points.The Society of Thoracic Surgeons (STS) registry data elements from 2086 isolated CABG patients were divided into training and testing datasets and input into Extreme Gradient Boosting decision-tree machine learning algorithms. Two prediction models were developed based on data from preoperative (80 parameters) and postoperative (125 parameters) phases of care. Outcomes included operative mortality, major morbidity or mortality, high cost, and 30-day readmission. Machine learning and STS model performance were assessed using accuracy and the area under the precision-recall curve (AUC-PR).Preoperative machine learning models predicted mortality (accuracy, 98%; AUC-PR = 0.16; F1 = 0.24), major morbidity or mortality (accuracy, 75%; AUC-PR = 0.33; F1 = 0.42), high cost (accuracy, 83%; AUC-PR = 0.51; F1 = 0.52), and 30-day readmission (accuracy, 70%; AUC-PR = 0.47; F1 = 0.49) with high accuracy. Preoperative machine learning models performed similarly to the STS for prediction of mortality (STS AUC-PR = 0.11; P = .409) and outperformed STS for prediction of mortality or major morbidity (STS AUC-PR = 0.28; P.001). Addition of intraoperative parameters further improved machine learning model performance for major morbidity or mortality (AUC-PR = 0.39; P.01) and high cost (AUC-PR = 0.64; P.01), with cross-clamp and bypass times emerging as important additive predictive parameters.Machine learning can predict mortality, major morbidity, high cost, and readmission after isolated CABG. Prediction based on the phase of care allows for dynamic risk assessment through the hospital course, which may benefit quality assessment and clinical decision-making.

In this article, we tackle the problem of predicting the “next” geographical position of a tourist, given her history (i.e., the prediction is done accordingly to the tourist’s current trail) by means of supervised learning techniques, namely Gradient Boosted Regression Trees and Ranking SVM. The learning is done on the basis of an object space represented by a 68-dimension feature vector specifically designed for tourism-related data. Furthermore, we propose a thorough comparison of several methods that are considered state-of-the-art in recommender and trail prediction systems for tourism, as well as a popularity baseline. Experiments show that the methods we propose consistently outperform the baselines and provide strong evidence of the performance and robustness of our solutions.

In the research of mining software repositories, we need to label a large amount of data to construct a predictive model. The correctness of the labels will affect the performance of a model substantially. However, limited studies have been performed to investigate the impact of mislabeled instances on a predictive model. To bridge the gap, in this work, we perform a case study on the security bug report (SBR) prediction. We found five publicly available datasets for SBR prediction contains many mislabeled instances, which lead to the poor performance of SBR prediction models of recent studies (e.g., the work of Peters et al. and Shu et al.). Furthermore, it might mislead the research direction of SBR prediction. In this paper, we first improve the label correctness of these five datasets by manually analyzing each bug report, and we find 749 SBRs, which are originally mislabeled as Non-SBRs (NSBRs). We then evaluate the impacts of datasets label correctness by comparing the performance of the classification models on both the noisy (i.e., before our correction) and the clean (i.e., after our correction) datasets. The results show that the cleaned datasets result in improvement in the performance of classification models. The performance of the approaches proposed by Peters et al. and Shu et al. on the clean datasets is much better than on the noisy datasets. Furthermore, with the clean datasets, the simple text classification models could significantly outperform the security keywords-matrix-based approaches applied by Peters et al. and Shu et al.

In the free online game industry, churn prediction is an important research topic. Reducing the churn rate of a game significantly helps with the success of the game. Churn prediction helps a game operator identify possible churning players and keep them engaged in the game via appropriate operational strategies, marketing strategies, and/or incentives. Playtime related features are some of the widely used universal features for most churn prediction models. In this paper, we consider developing new universal features for churn predictions for long-term players based on players' playtime. In particular, we measure playtime regularity using the notion of entropy and cross-entropy from information theory. After we calculate the playtime regularity of players from data sets of six free online games of different types. We leverage information from players' playtime regularity in the form of universal features for churn prediction. Experiments show that our developed features are better at predicting churners compared to baseline features. Thus, the experiment results imply that our proposed features could utilize the information extracted from players' playtime more effectively than related baseline playtime features.

Ozone (O3) is one of the common air pollutants. An increase in the ozone concentration can adversely affect public health and the environment such as vegetation and crops. Therefore, atmospheric air quality monitoring systems were found to monitor and predict ozone concentration. Due to complex formation of ozone influenced by precursors of ozone (O3) and meteorological conditions, there is a need to examine and evaluate various machine learning (ML) models for ozone concentration prediction. This study aims to utilize various ML models including Linear Regression (LR), Tree Regression (TR), Support Vector Regression (SVR), Ensemble Regression (ER), Gaussian Process Regression (GPR) and Artificial Neural Networks Models (ANN) to predict tropospheric (O3) using ozone concentration dataset. The dataset was created by observing hourly average data from air quality monitoring systems in 3 different stations including Putrajaya, Kelang, and KL in 3 sites in Peninsular Malaysia. The prediction models have been trained on this dataset and validated by optimizing their hyperparameters. Additionally, the performance of models was evaluated in terms of RMSE, MAE, R2, and training time. The results indicated that LR, SVR, GPR and ANN were able to give the highest R2 (83 % and 89 %) with specific hyperparameters in stations Kelang and KL, respectively. On the other hand, SVR and ER outweigh other models in terms of R2 (79 %) in Putrajaya station. Overall, regardless slightly performance differences, several developed models were able to learn patterns well and provide good prediction performance in terms of R2, RMSE and MAE. Ensemble regression models were found to balance between high prediction accuracy in terms of R2 and low training time and thus considered as a feasible solution for application of Ozone concentration prediction using the data in hourly scenario.

This paper proposes a customized kernel-based Fuzzy C-Means (CKFCM) clustering, which provides an original framework for data interpretation and data analysis and delivers an effective solution for an accurate division of actual run data under multiple operating conditions in an iron ore sintering process (IOSP). The improvement of CKFCM clustering is achieved by including an adjustment factor introduced into the kernel-based Fuzzy C-Means clustering. The adjustment factor only needs to consider a small amount of labeled data that are determined based on expert experience. Subsequently, the CKFCM clustering is applied to the modeling of the IOSP, and Spearman's rank correlation coefficient method is utilized to determine input variables of the model under different operating conditions. Furthermore, the broad learning model is employed to build the prediction model for each operating condition. Finally, we conducted an in-depth analysis of the presented clustering method. Its performance has been experimentally demonstrated by many publicly available datasets. Meanwhile, simulation results involving actual run data of the IOSP demonstrate the superiority and effectiveness of the developed model in carbon efficiency prediction. We show that the developed model outperforms the state-of-the-art prediction models in achieving a good balance between simplicity and accuracy.

Heart disease seriously threatens human life due to high morbidity and mortality. Accurate prediction and diagnosis become more critical for early prevention, detection and treatment. The Internet of Medical Things (IoMT) and Artificial Intelligence (AI) support healthcare services in heart disease monitoring, prediction and diagnosis. However, most prediction models only predict whether people are sick, and rarely further determine the severity of the disease. In this paper, we propose a machine learning based prediction model to achieve binary and multiple classification heart disease prediction simultaneously. We first design a Fuzzy-GBDT algorithm combining fuzzy logic and Gradient Boosting Decision Tree (GBDT) to reduce data complexity and increase the generalization of binary classification prediction. Then, we integrate Fuzzy-GBDT with Bagging to avoid over-fitting. The Bagging-Fuzzy-GBDT for multi-classification prediction further classify the severity of heart disease. Evaluation results demonstrate the Bagging-Fuzzy-GBDT has excellent accuracy and stability in both binary and multiple classification predictions.

Time-series data are widespread in real-world industrial scenarios. To recover and infer missing information in real-world applications, the problem of time-series prediction has been widely studied as a classical research topic in data mining. Deep learning architectures have been viewed as next-generation time-series prediction models. However, recent studies have shown that deep learning models are vulnerable to adversarial attacks. In this study, we prospectively examine the problem of time-series prediction adversarial attacks and propose an attack strategy for generating an adversarial time series by adding malicious perturbations to the original time series to deteriorate the performance of time-series prediction models. Specifically, a perturbation-based adversarial example generation algorithm is proposed using the gradient information of the prediction model. In practice, unlike the imperceptibility to humans in the field of image processing, time-series data are more sensitive to abnormal perturbations and there are more stringent requirements regarding the amount of perturbations. To address this challenge, we craft an adversarial time series based on the importance measurement to slightly perturb the original data. Based on comprehensive experiments conducted on real-world time-series datasets, we verify that the proposed adversarial attack methods not only effectively fool the target time-series prediction model LSTNet, they also attack state-of-the-art CNN-, RNN-, and MHANET-based models. Meanwhile, the results show that the proposed methods achieve a good transferability. That is, the adversarial examples generated for a specific prediction model can significantly affect the performance of the other methods. Moreover, through a comparison with existing adversarial attack approaches, we can see that much smaller perturbations are sufficient for the proposed importance-measurement based adversarial attack method. The methods described in this paper are significant in understanding the impact of adversarial attacks on a time-series prediction and promoting the robustness of such prediction technologies.

Cross-project defect prediction (CPDP), aiming to apply defect prediction models built on source projects to a target project, has been an active research topic. A variety of supervised CPDP methods and some simple unsupervised CPDP methods have been proposed. In a recent study, Zhou et al. found that simple unsupervised CPDP methods (i.e., ManualDown and ManualUp) have a prediction performance comparable or even superior to complex supervised CPDP methods. Therefore, they suggested that the ManualDown should be treated as the baseline when considering non-effort-aware performance measures (NPMs) and the ManualUp should be treated as the baseline when considering effort-aware performance measures (EPMs) in future CPDP studies. However, in that work, these unsupervised methods are only compared with existing supervised CPDP methods in terms of one or two NPMs and the prediction results of baselines are directly collected from the primary literature. Besides, the comparison has not considered other recently proposed EPMs, which consider context switches and developer fatigue due to initial false alarms. These limitations may not give a holistic comparison between the supervised methods and unsupervised methods. In this paper, we aim to revisit Zhou et al.'s study. To the best of our knowledge, we are the first to make a comparison between the existing supervised CPDP methods and the unsupervised methods proposed by Zhou et al. in the same experimental setting, considering both NPMs and EPMs. We also propose an improved supervised CPDP method EASC and make a further comparison between this method and the unsupervised methods. According to the results on 82 projects in terms of 12 performance measures, we find that when considering NPMs, EASC can achieve similar results with the unsupervised method ManualDown without statistically significant difference in most cases. However, when considering EPMs, our proposed supervised method EASC can statistically significantly outperform the unsupervised method ManualUp with a large improvement in terms of Cliff's delta in most cases. Therefore, the supervised CPDP methods are more promising than the unsupervised method in practical application scenarios, since the limitation of testing resource and the impact on developers cannot be ignored in these scenarios.

Traffic prediction is a spatiotemporal predictive task that plays an essential role in intelligent transportation systems. Today, graph convolutional neural networks (GCNNs) have become the prevailing models in the traffic prediction literature since they excel at extracting spatial correlations. In this work, we classify the components of successful GCNN prediction models and analyze the effects of matrix factorization, attention mechanism, and weight sharing on their performance. Furthermore, we compare these variations against random forests, a traditional regression method that predates GCNNs by over 15 years. We evaluated these methods using simulated data of two regions in Toronto as well as real-world sensor data from selected California highways. We found that incorporating matrix factorization, attention, and location-specific model weights either individually or collectively into GCNNs can result in a better overall performance. Moreover, although random forest regression is a less compact model, it matches or exceeds the performance of all variations of GCNNs in our experiments. This suggests that the current graph convolutional methods may not be the best approach to traffic prediction and there is still room for improvement. Finally, our findings also suggest that for future research on GCNN for traffic prediction to be credible, researchers must include performance comparison to random forests.

Accurate diagnosis of a disease is essential in healthcare. Prediction models, based on classical regression techniques, are widely used in clinical practice. Machine Learning (ML) techniques might be preferred in case of a large amount of data per patient and relatively limited numbers of subjects. However, this increases the risk of overfitting, and external validation is imperative. However, in the field of ML, new and more efficient techniques are developed rapidly, and if recruiting patients for a validation study is time consuming, the ML technique used to develop the first model might have been surpassed by more efficient ML techniques, rendering this original model no longer relevant. We demonstrate a stepwise design for simultaneous development and validation of prediction models based on ML techniques. The design enables - in one study - evaluation of the stability and robustness of a prediction model over increasing sample size as well as assessment of the stability of sensitivity/specificity at a chosen cut-off. This will shorten the time to introduction of a new test in health care. We finally describe how to use regular clinical parameters in conjunction with ML based predictions, to further enhance differentiation between subjects with and without a disease.

Accumulated studies have discovered that circular RNAs (CircRNAs) are closely related to many complex human diseases. Due to this close relationship, CircRNAs can be used as good biomarkers for disease diagnosis and therapeutic targets for treatments. However, the number of experimentally verified circRNA-disease associations are still fewer and also conducting wet-lab experiments are constrained by the small scale and cost of time and labour. Therefore, effective computational methods are required to predict associations between circRNAs and diseases which will be promising candidates for small scale biological and clinical experiments. In this paper, we propose novel computational models based on Graph Convolution Networks (GCN) for the potential circRNA-disease association prediction. Currently most of the existing prediction methods use shallow learning algorithms. Instead, the proposed models combine the strengths of deep learning and graphs for the computation. First, they integrate multi-source similarity information into the association network. Next, models predict potential associations using graph convolution which explore this important relational knowledge of that network structure. Two circRNA-disease association prediction models, GCN based Node Classification (GCN-NC) and GCN based Link Prediction (GCN-LP) are introduced in this work and they demonstrate promising results in various experiments and outperforms other existing methods. Further, a case study proves that some of the predicted results of the novel computational models were confirmed by published literature and all top results could be verified using gene-gene interaction networks.

Forecasting solar irradiance is of utmost importance in supplying renewable energy efficiently and timely. This paper aims to experiment five variants of recurrent neural networks (RNN), and develop effective and reliable 5-minute short term solar irradiance prediction models. The 5 RNN classes are long-short term memory (LSTM), gated recurrent unit (GRU), Simple RNN, bidirectional LSTM (Bi-LSTM), and bidirectional GRU (Bi-GRU); the first 3 classes are unidirectional and the last two are bidirectional RNN models. The 26 months data under consideration, exhibits extremely volatile weather conditions in Jinju city, South Korea. Therefore, after different experimental processes, 5 hyper-parameters were selected for each model cautiously. In each model, different levels of depth and width were tested; moreover, a 9-fold cross validation was applied to distinguish them against high variability in the seasonal time-series dataset. Generally the deeper architectures of the aforementioned models had significant outcomes; meanwhile, the Bi-LSTM and Bi-GRU provided more accurate predictions as compared to the unidirectional ones. The Bi-GRU model provided the lowest RMSE and highest R2 values of 46.1 and 0.958; additionally, it required 5.25*10-5 seconds per trainable parameter per epoch, the lowest incurred computational cost among the mentioned models. All 5 models performed differently over the four seasons in the 9-fold cross validation test. On average, the bidirectional RNNs and the simple RNN model showed high robustness with less data and high temporal data variability; although, the stronger architectures of the bidirectional models, deems their results more reliable.

To address the problem that the high dimensionality, time-series, coupling and multiple time scales of production data in the process industry lead to the low accuracy of traditional prediction models, we propose a multi-time scale data enhancement and cement clinker f-CaO prediction method based on Regression-Wasserstein Generative Adversarial Nets (R-WGAN) model. The model is built using a combination of WGAN and regression prediction networks. Firstly, the data is extracted according to the principle of Sliding Window to eliminate the effect of time-varying delay between data in data enhancement and prediction, and a Dual Data Pathway is used for data stitching so that data of different time scales can be enhanced at the same time. We then augment the data with a generator network, use a discriminator network to judge the goodness of the generated data, and propose an auxiliary evaluation strategy to evaluate whether the high-dimensional generated data conform to the actual laws, expand the training set of the regression prediction network with the generated data that conform to the laws, and finally achieve the prediction of cement clinker f-CaO. The model was applied in the quality management system of a Cement Company for simulation, and experiments showed that the model with data enhancement has the advantages of high accuracy, robustness and good generalization in cement clinker f-CaO prediction.

Stroke has become a leading cause of death and long-term disability in the world, and there is no effective treatment.Deep learning-based approaches have the potential to outperform existing stroke risk prediction models, they rely on large well-labeled data. Due to the strict privacy protection policy in health-care systems, stroke data is usually distributed among different hospitals in small pieces. In addition, the positive and negative instances of such data are extremely imbalanced. Transfer learning solves small data issue by exploiting the knowledge of a correlated domain, especially when multiple source are available.In this work, we propose a novel Hybrid Deep Transfer Learning-based Stroke Risk Prediction (HDTL-SRP) scheme to exploit the knowledge structure from multiple correlated sources (i.e.,external stroke data, chronic diseases data, such as hypertension and diabetes). The proposed framework has been extensively tested in synthetic and real-world scenarios, and it outperforms the state-of-the-art stroke risk prediction models. It also shows the potential of real-world deployment among multiple hospitals aided with 5G/B5G infrastructures. National Key R&D Program of China; National Nature Science Foundation of China; Natural Science Foundation of Guangdong Province; Guangdong “Pearl River Talent Recruitment Program”; Technology Research Project of Shenzhen City; Public Technology Platform of Shenzhen City

As a powerful tool for data streams processing, the vast majority of existing evolving intelligent systems (EISs) learn prediction models from data in a supervised manner. However, high-quality labelled data can be difficult to obtain in many real-world classification applications concerning data streams, though unlabelled data is plentiful. To overcome the labelling bottleneck and construct a stronger classification model, a novel semi-supervised EIS is proposed in this paper. After being primed with a small amount of labelled data, the proposed method is capable of continuously self-developing its system structure and self-updating the meta-parameters from unlabelled data streams chunk-by-chunk in a non-iterative, exploratory manner by exploiting a novel pseudo-labelling strategy. Thanks to its transparent prototype-based structure and human-understandable reasoning process, the proposed method can provide users high explainability and interpretability while achieving great classification precision. Experimental investigation demonstrates the superior performance of the proposed method.

Accurate wind power forecasting is a vital factor in day-ahead dispatch and increasing the level of penetration of renewable energy. The feature extraction of meteorological factors related to wind power output is a big challenge to improve forecasting accuracy, and selecting key meteorological factors based on experience will decrease the prediction accuracy. Therefore, a day-ahead wind power combined forecasting approach is innovatively proposed through key meteorological factors selection, data decomposition and reconstruction, combined forecasting model generation, and optimization strategy. Correlated variables namely variational mode decomposition and weighted permutation entropy (VMD-WPE) decomposed historical wind power and key meteorological factors are used as the inputs. A forecasting model based on convolutional neural network (CNN) and long short-term memory network (LSTM) is used to forecast future wind power. Four optimizers with different optimization performances are used to find the best parameters of the forecasting model to obtain accurate prediction results. Multiple comparative experiments from regional wind farms in Ningxia and Jilin of China are utilized as case studies to evaluate the effectiveness of the proposed model. Results show that the proposed approach outperforms other benchmark prediction models, taking into account multiple-error metrics including error metrics, accuracy rate, and improvement percentages.

The quantitative structure-activity relationship (QSAR) approach has been used in numerous chemical compounds as in silico computational assessment for a long time. Further, owing to the high-performance modeling of QSAR, machine learning methods have been developed and upgraded. Particularly, the three- dimensional structure of chemical compounds has been gaining increasing attention owing to the representation of a large amount of information. However, only many of feature extraction is impossible to build models with the high-ability of the prediction. Thus, suitable extraction and effective selection of features are essential for models with excellent performance. Recently, the deep learning method has been employed to construct prediction models with very high performance using big data, especially, in the field of classification. Therefore, in this study, we developed a molecular image-based novel QSAR approach, called DeepSnap-Deep learning approach for designing high-performance models. In addition, this DeepSnap-Deep learning approach outperformed the conventional machine learnings when they are compared. Herein, we discuss the advantage and disadvantages of the machine learnings as well as the availability of the DeepSnap-Deep learning approach.

Eutrophication and excessive algal growth pose a threat on aquatic organisms and the health of the public, environment, and the economy. Understanding what drives excessive algal growth can inform mitigation measures and aid in advance planning to minimize impacts. We demonstrate how simulated data from weather, hydrological, and agroecosystem numerical prediction models can be combined with machine learning (ML) to assess and predict chlorophyll a (chl a) concentrations, a proxy for lake eutrophication and algal biomass. The study area is Lake Erie for a 16-year period, 2002–2017. A total of 20 environmental variables from linked and coupled physical models are used as input features to train the ML model with chl a observations from 16 measuring stations. Included are meteorological variables from the Weather Research and Forecasting (WRF) model, hydrological variables from the Variable Infiltration Capacity (VIC) model, and agricultural management practice variables from the Environmental Policy Integrated Climate (EPIC) agroecosystem model. The consolidation of these variables is conducive to a successful prediction of chl a. Aside from the synergistic effects that weather, hydrology, and fertilizers have on eutrophication and excessive algal growth, we found that the application of different forms of both P and N fertilizers are highly ranked for the prediction of chl a concentration. The developed ML model successfully predicts chl a with a coefficient of determination of 0.81, bias of −0.12 μg/l and RMSE of 4.97 μg/l. The developed ML-based modeling approach can be used for impact assessment of agriculture practices in a changing climate that affect chl a concentrations in Lake Erie.

Objectives: No clear guidance exists on handling missing data at each stage of developing, validating and implementing a clinical prediction model (CPM). We aimed to review the approaches to handling missing data that underly the CPMs currently recommended for use in UK healthcare.Study Design and Setting: A descriptive cross-sectional meta-epidemiological study aiming to identify CPMs recommended by the National Institute for Health and Care Excellence (NICE), which summarized how missing data is handled across their pipelines.Results: A total of 23 CPMs were included through “sampling strategy.” Six missing data strategies were identified: complete case analysis (CCA), multiple imputation, imputation of mean values, k-nearest neighbours imputation, using an additional category for missingness, considering missing values as risk-factor-absent. 52% of the development articles and 48% of the validation articles did not report how missing data were handled. CCA was the most common approach used for development (40%) and validation (44%). At implementation, 57% of the CPMs required complete data entry, whilst 43% allowed missing values. Three CPMs had consistent paths in their pipelines.Conclusion: A broad variety of methods for handling missing data underly the CPMs currently recommended for use in UK healthcare. Missing data handling strategies were generally inconsistent. Better quality assurance of CPMs needs greater clarity and consistency in handling of missing data.

Flow-accelerated corrosion (FAC) of carbon steel piping is a significant problem in nuclear power plants. The basic process of FAC is currently understood relatively well; however, the accuracy of prediction models of the wall-thinning rate under an FAC environment is not reliable. Herein, we propose a methodology to construct pipe wall-thinning rate prediction models using artificial neural networks and a convolutional neural network, which is confined to a straight pipe without geometric changes. Furthermore, a methodology to generate training data is proposed to efficiently train the neural network for the development of a machine learning-based FAC prediction model. Consequently, it is concluded that machine learning can be used to construct pipe wall thinning rate prediction models and optimize the number of training datasets for training the machine learning algorithm. The proposed methodology can be applied to efficiently generate a large dataset from an FAC test to develop a wall thinning rate prediction model for a real situation.

BACKGROUND Research on predictors and risk of recurrence after suicide attempt from China is lacking. This study aims to identify risk factors and develop prediction models for recurrent suicidal behavior among suicide attempters using Cox proportional hazard (CPH) and machine learning methods. METHODS The prospective cohort study included 1103 suicide attempters with a maximum follow-up of 10 years from rural China. Baseline characteristics, collected by face-to-face interviews at least 1 month later after index suicide attempt, were used to predict recurrent suicidal behavior. CPH and 3 machine learning algorithms, namely, the least absolute shrinkage and selection operator, random survival forest, and gradient boosting decision tree, were used to construct prediction models. Model performance was accessed by concordance index (C-index) and the time-dependent area under the receiver operating characteristic curve (AUC) value for discrimination, and time-dependent calibration curve along with Brier score for calibration. RESULTS The median follow-up time was 7.79 years, and 49 suicide attempters had recurrent suicidal behavior during the study period. Four models achieved comparably good discrimination and calibration performance, with all C-indexes larger than 0.70, AUC values larger than 0.65, and Brier scores smaller than 0.06. Mental disorder emerged as the most important predictor across all four models. Suicide attempters with mental disorders had a 3 times higher risk of recurrence than those without. History of suicide attempt (HR = 2.84, 95% CI: 1.34-6.02), unstable marital status (HR = 2.81, 95% CI: 1.38-5.71), and older age (HR = 1.51, 95% CI: 1.14-2.01) were also identified as independent predictors of recurrent suicidal behavior by CPH model. CONCLUSIONS We developed four models to predict recurrent suicidal behavior with comparable good prediction performance. Our findings potentially provided benefits in screening vulnerable individuals on a more precise scale.

Customer churn prediction models using machine learning classification have been developed predominantly by training and testing on one time slice of data. We train models on multiple time slices of data and refer to this approach as multi-slicing. Our results show that given the same time frame of data, multi-slicing significantly improves churn prediction performance compared to training on the entire data set as one time slice. We demonstrate that besides an increased training set size, the improvement is driven by training on samples from different time slices. For data from a convenience wholesaler, we show that multi-slicing addresses the rarity of churn samples and the risk of overfitting to the distinctive situation in a single training time slice. Multi-slicing makes a model more generalizable, which is particularly relevant whenever conditions change or fluctuate over time. We also discuss how to choose the number of time slices.

The scale of criminal networks (e.g. drug syndicates and terrorist networks) extends globally and poses national security threat to many nations as they also tend to be technologically advance (e.g. Dark Web and Silk Road cryptocurrency). Therefore, it is critical for law enforcement agencies to be equipped with the latest tools in criminal network analysis (CNA) to obtain key hidden links (relationships) within criminal networks to preempt and disrupt criminal network structures and activities. Current hidden or missing link predictive models that are based on Social Network Analysis models rely on ML techniques to improve the performance of the models in terms of predictive accuracy and computing power. Given the improvement in the recent performance of Deep Reinforcement Learning (DRL) techniques which could train ML models through self-generated dataset, DRL can be usefully applied to domains with relatively smaller dataset such as criminal networks. The objective of this study is to assess the comparative performance of a CNA hidden link prediction model developed using DRL techniques against classical ML models such as gradient boosting machine (GBM), random forest (RF) and support vector machine (SVM). The experiment results exhibit an improvement in the performance of the DRL model of about 7.4% over the next best performing classical RF model trained within 1500 iterations. The performance of these link prediction models can be scaled up with the parallel processing capabilities of graphical processing units (GPUs), to significantly improve the speed of training the model and the prediction of hidden links.

To achieve a high penetration of renewable energy integration, an effective solution is to explore the interdependence between numerical weather prediction (NWP) data and historical wind power to improve prediction accuracy. This paper proposes a novel combined approach for wind power prediction. The characteristics of NWP and historical wind power data are extracted by using the feature extraction technique, the predictor is designed based on extreme learning machine (ELM) and least squares support vector machine (LSSVM) model, and then key parameters of the prediction models are optimized by improving cuckoo search (ICS) to obtain a reliable value, which is defined as the pre-combined prediction value (PPA). To obtain a reliable result, a variance strategy is developed to allocate the weights of the pre-combined prediction model to obtain the final predicted values. Four seasons dataset collected from regional wind farms in China is utilized as a benchmark experiment to evaluate the effectiveness of the proposed approach. The results of comprehensive numerical cases with different seasons show that the proposed approach, which considers multiple-error metrics, including error metrics, accuracy rate, qualification rate, and improvement percentages, achieves higher accuracy than other benchmark prediction models.

Stock price prediction is important in both financial and commercial domains, and using neural networks to forecast stock prices has been a topic of ongoing research and development. Traditional prediction models are often based on a single type of data and do not account for the interplay of many variables. This study covers a radial basis neural network modeling technique with multiview collaborative learning capabilities for incorporating the impacts of numerous elements into the prediction model. This research offers a multiview RBF neural network prediction model based on the classic RBF network by integrating a collaborative learning item with multiview learning capabilities (MV-RBF). MV-RBF can make full use of both the internal information provided by the correlation between each view and the distinct characteristics of each view to form independent sample information. By using two separate stock qualities as input feature information for trials, this study proves the viability of the multiview RBF neural network prediction model on a real data set.

In Bangladesh, groundwater is the main source of both drinking water and irrigation. Suction lift pumps and force mode of operation are the predominant technologies for groundwater abstraction in Bangladesh. For a sustainable usage policy, it is thus important to identify which technology would be more appropriate for which area in Bangladesh. With that aim in mind, this paper proposes a methodology that leverages the power of machine learning that can potentially learn intricate relationships between the (annual maximum) groundwater level (GWL) and the relevant hydrogeological factors (HGFs). A number of machine learning algorithms—both classification and regression models—was trained. Our classification models were trained as a binary classifier to predict the abstraction technology of a particular point. Notably, our best classification model was based on the Random Forest algorithm, which achieved an accuracy of 91% and an excellent value of 96% for the area under receiver operating characteristics curve, indicating its strong discriminant capability. We also identified (elevation derived from) digital elevation model, specific yield and lithology as the three most important HGFs for GWL in Bangladesh. On the other hand, to predict the actual (annual maximum) GWL, we employed a two-stage approach, where we first employed the above-mentioned classification model to identify the suitable abstraction technology for the point of interest and subsequently predict the actual GWL using the appropriate Random Forest regressor. This also had a reasonable accuracy (minimum absolute error was less than 1 for suction mode and less than 5 for the force mode). Finally, using our prediction models, we prepared groundwater (technology) maps for the whole Bangladesh.

The first case of COVID-19 in Eswatini was first reported in March 2020, posing an unprecedented challenge to the country’s health and socio-economic systems. Using geographic information system (GIS) data comprising 15 socioeconomic, demographic and environmental variables, we model the spatial variability of COVID-19 transmission risk based on case data for the period under strict lockdown (up to 8th May 2020) and after the lockdown regulations were gradually eased (up to 30th June 2020). We implemented and tested 13 spatial data-driven Bayesian network (BN) learning algorithms to examine the factors that determine the spatial distribution of COVID-19 transmission risk. All the BN models performed very well in predicting the COVID-19 cases as evidenced by low log loss (0.705–0.683) and high recall values (0.821–0.836). The tree-augmented naïve (TAN) model outperformed all other BN learning algorithms. The proximity to major health facilities, churches, shopping centres and supermarkets as well as average annual traffic density were the strongest predictors of transmission risk during strict lockdown. After gradual relaxation of the lockdown, the proportion of the youth (15–40 years old) in an area became the strongest predictor of COVID-19 transmission in addition to the proximity to areas where people congregate, excluding churches. The study provides useful insights on the spatio-temporal dynamics of COVID-19 transmission drivers thereby aiding the design of geographically-targeted interventions. The findings also point to the robustness of BN models in spatial predictive modelling and graphically explaining spatial phenomena under uncertainty and with limited data.

This study investigates the contribution of Boreal Summer Intraseasonal Oscillation (BSISO) to the tropical cyclone (TC) activity over the North Indian Ocean (NIO) and assesses the prediction skill of a statistical Generalised Additive Model (GAM) and two machine learning techniques—Random Forest (RF) and Support Vector Regression (SVR). Joint Typhoon Warning Centre TC and BSISO1 Index data have been used for a period of 33-year (1981–2013). By considering eight phases of BSISO, prediction models have been developed using a kernel density estimation for the TC genesis, Euler integration step to fit the tracks, and a country mask approach for the landfall across the NIO rim countries. Result shows that GAM has the highest prediction skill compared to the RF and SVR. Westward and Northward moving TCs are controlled by the wind and the TC activities during BSISO phases which modulated by wind matched well against observations over the NIO. Distance calculation validation method is applied to assess the skill of models.

Objective This article reports a systematic review of studies containing development and validation of models predicting postacute care destination after adult inpatient hospitalization, summarizes clinical populations and variables, evaluates model performance, assesses risk of bias and applicability, and makes recommendations to reduce bias in future models. Materials and Methods A systematic literature review was conducted following PRISMA guidelines and the Cochrane Prognosis Methods Group criteria. Online databases were searched in June 2020 to identify all published studies in this area. Data were extracted based on the CHARMS checklist, and studies were evaluated based on predictor variables, validation, performance in validation, risk of bias, and applicability using the Prediction Model Risk of Bias Assessment Tool (PROBAST) tool. Results The final sample contained 28 articles with 35 models for evaluation. Models focused on surgical (22), medical (5), or both (8) populations. Eighteen models were internally validated, 10 were externally validated, and 7 models underwent both types. Model performance varied within and across populations. Most models used retrospective data, the median number of predictors was 8.5, and most models demonstrated risk of bias. Discussion and Conclusion Prediction modeling studies for postacute care destinations are becoming more prolific in the literature, but model development and validation strategies are inconsistent, and performance is variable. Most models are developed using regression, but machine learning methods are increasing in frequency. Future studies should ensure the rigorous variable selection and follow TRIPOD guidelines. Only 14% of the models have been tested or implemented beyond original studies, so translation into practice requires further investigation.

In chemical processes, grasping the changing trend of critical parameters can help field operators take appropriate adjustments to eliminate potential fluctuations. Thus, deep networks, renowned for its revolutionary feature representation capability, have been gradually exploited for building reliable prediction models from massive data embraced tremendously nonlinearities and dynamics. Because of the inherent complexity, the process trajectories over the whole running duration make distinctive contributions to the ultimate targets. Specifically, features extracted from different secondary variables at different previous instants have diverse impacts on the current state of primary variables. However, this spatiotemporal relevance discrepancy is rarely considered, which may lead to deterioration of prediction performance. Therefore, this paper seamlessly integrates the spatiotemporal attention (STA) mechanism with convolutional neural networks (CNN) and bi-directional long short-term memory (BiLSTM), and proposes a novel predictive model, namely STA-ConvBiLSTM. Using the deep framework composed of CNN and BiLSTM, the integrated model can, not only automatically explore the esoteric spatial correlations among high-dimensional variables at each time step, but also adaptively excavate beneficial temporal characteristics across all time steps. Meanwhile, STA is further introduced to assign corresponding weights to information with dissimilar importance, so as to prevent high target-relevant interactions from being discarded due to overlong sequences and excessive features. STA-ConvBiLSTM is applied in the case of furnace tube temperature prediction of a delayed coking unit, which exhibits a significant improvement of the prediction accuracy.

As Photovoltaic (PV) energy is impacted by various weather variables such as solar radiation and temperature, one of the key challenges facing solar energy forecasting is choosing the right inputs to achieve the most accurate prediction. Weather datasets, past power data sets, or both sets can be utilized to build different forecasting models. However, operators of grid-connected PV farms do not always have full sets of data available to them especially over an extended period of time as required by key techniques such as multiple regression (MR) or artificial neural network (ANN). Therefore, the research reported here considered these two main approaches of building prediction models and compared their performance when utilizing structural, time-series, and hybrid methods for data input. Three years of PV power generation data (of an actual farm) as well as historical weather data (of the same location) with several key variables were collected and utilized to build and test six prediction models. Models were built and designed to forecast the PV power for a 24-hour ahead horizon with 15 min resolutions. Results of comparative performance analysis show that different models have different prediction accuracy depending on the input method used to build the model: ANN models perform better than the MR regardless of the input method used. The hybrid input method results in better prediction accuracy for both MR and ANN techniques, while using the time-series method results in the least accurate forecasting models. Furthermore, sensitivity analysis shows that poor data quality does impact forecasting accuracy negatively especially for the structural approach.

Many conventional quality prediction models are directly developed based on the easy-to-measure variables, and thus the local information within individual unit may be buried by information of other units. In this study, a cascaded regression network (RegNet) is proposed to solve the aforementioned issue. Specifically, the features which are adopted to develop RegNet model are extracted in two steps, including variable-wise and unit-wise feature extractions. In variable-wise feature extraction, several adjacent variables and their corresponding time lags are integrated using convolutional filter. By this means, both local correlation and temporal information within each unit can be preserved. In the unit-wise feature extraction, the local information of each unit is adopted to further explore the global correlation between different operation units. Based on the obtained global features, a fully connected layer is designed to calculate the regression weight of the quality prediction model. It is noted that the architecture of RegNet can be readily generalized to many existing methods by replacing the convolutional filter and fully connected layer. The performance of the proposed method is illustrated using a simulated process and two real industrial processes, and the experimental results show that it can provide reliable prediction results for industrial applications.

Objective There are growing interests on suicide risk screening in clinical settings and classifying high-risk groups of suicide with suicidal ideation is crucial for a more effective suicide preventive intervention. Previous statistical techniques were limited because they tried to predict suicide risk using a simple algorithm. Machine learning differs from the traditional statistical techniques in that it generates the most optimal algorithm from various predictors.Methods We aim to analyze the Personality Assessment Inventory (PAI) profiles of child and adolescent patients who received outpatient psychiatric care using machine learning techniques, such as logistic regression (LR), random forest (RF), artificial neural network (ANN), support vector machine (SVM), and extreme gradient boosting (XGB), to develop and validate a classification model for individuals with high suicide risk.Results We developed prediction models using seven relevant features calculated by Boruta algorithm and subsequently tested all models using the testing dataset. The area under the ROC curve of these models were above 0.9 and the RF model exhibited the best performance.Conclusion Suicide must be assessed based on multiple aspects, and although Personality Assessment Inventory for Adolescent assess an array of domains, further research is needed for predicting high suicide risk groups.

Solubility has been widely regarded as a fundamental property of small molecule drugs and drug candidates, as it has a profound impact on the crystallization process. Solubility prediction, as an alternative to experiments which can reduce waste and improve crystallization process efficiency, has attracted increasing attention. However, there are still many urgent challenges thus far. Herein we used seven descriptors based on understanding dissolution behavior to establish two solubility prediction models by machine learning algorithms. The solubility data of 120 active pharmaceutical ingredients (APIs) in ethanol were considered in the prediction models, which were constructed by random decision forests and artificial neural network with optimized data structure and model accuracy. Furthermore, a comparison with traditional prediction methods including the modified solubility equation and the quantitative structure-property relationships model was carried out. The highest accuracy shown by the testing set proves that the ML models have the best solubility prediction ability. Multiple linear regression and stepwise regression were used to further investigate the critical factor in determining solubility value. The results revealed that the API properties and the solute-solvent interaction both provide a nonnegligible contribution to the solubility value.

Objective Low-density lipoprotein cholesterol (LDL-C) can be estimated using the Friedewald and Martin-Hopkins formulas. We developed LDL-C prediction models using multiple machine learning methods and investigated the validity of the new models along with the former formulas. Methods Laboratory data (n = 59,415) on measured LDL-C, high-density lipoprotein cholesterol, triglycerides (TG), and total cholesterol were partitioned into training and test data sets. Linear regression, gradient-boosted trees, and artificial neural network (ANN) models were formed based on the training data. Paired-group comparisons were performed using a t-test and the Wilcoxon signed-rank test. We considered P values .2 to be statistically significant. Results For TG ≥177 mg/dL, the Friedewald formula underestimated and the Martin-Hopkins formula overestimated the LDL-C (P .001 vs. P Conclusion Linear regression, gradient-boosted trees, and ANN models offer more accurate alternatives to the aforementioned formulas, especially for TG 177 to 399 mg/dL and LDL-C

For the purpose of reducing the harm of expressway traffic accidents and improving the accuracy of traffic accident black spots identification, this paper proposes a method for black spots identification of expressway accidents based on road unit secondary division and empirical Bayes method. Based on the modelling ideas of expressway accident prediction models in HSM (Highway Safety Manual), an expressway accident prediction model is established as a prior distribution and combined with empirical Bayes method safety estimation to obtain a Bayes posterior estimate. The posterior estimated value is substituted into the quality control method to obtain the black spots identification threshold. Finally, combining the Xi'an-Baoji expressway related data and using the method proposed in this paper, a case study of Xibao Expressway is carried out, and sections 9, 19, and 25 of Xibao Expressway are identified as black spots. The results show that the method of secondary segmentation based on dynamic clustering can objectively describe the concentration and dispersion of accident spots on the expressway, and the proposed black point recognition method based on empirical Bayes method can accurately identify accident black spots. The research results of this paper can provide a basis for decision-making of expressway management departments, take targeted safety improvement measures.

It is not easy for banks to sell their term-deposit products to new clients because many factors will affect customers’ purchasing decision and because banks may have difficulties to identify their target customers. To address this issue, we use different supervised machine learning algorithms to predict if a customer will subscribe a bank term deposit and then compare the performance of these prediction models. Specifically, the current paper employs these five algorithms: Naïve Bayes, Decision Tree, Random Forest, Support Vector Machine and Neural Network. This paper thus contributes to the artificial intelligence and Big Data field with an important evidence of the best performed model for predicting bank term deposit subscription.

Accurate and reliable suspended sediment load (SSL) prediction models are necessary for the planning and management of water resource structures. In this study, four machine learning techniques, namely Gradient boost regression (GBT), Random Forest (RF), Support vector machine (SVM), and Artificial neural network ANN will be developed to predict SSL at the Rantau Panjang station on Johor River basin (JRB), Malaysia. Four evaluation criteria, including the Correlation Coefficient (R), Root Mean Square Error (RMSE), Nash Sutcliffe Efficiency (NSE) and Scatter Index (SI) will utilize to evaluating the performance of the proposed models. The obtained results revealed that all the proposed Machine Learning (ML) models showed superior prediction daily SSL performance. The comparative outcomes among models were carried out using the Taylor diagram. ANN model shows more reliable results than other models with R of 0.989, SI of 0.199, RMSE of 0.011053 and NSE of 0.979. A sensitivity analysis of the models to the input variables revealed that the absence of current day Suspended sediment load data SSLt-1 had the most effect on the SSL. Moreover, to examine validation of most accurate model we proposed divided data to 50% training, 25% testing and 25% validation) sets and ANN provided superior performance. Therefore, the proposed ANN approach is recommended as the most accurate model for SSL prediction.

Key messageWe propose a novel approach to the Bayesian optimization of multi-variate genomic prediction models based on secondary traits to improve accuracy gains and phenotyping costs via efficient Pareto frontier estimation.Multivariate genomic prediction based on secondary traits, such as data from various omics technologies including high-throughput phenotyping (e.g., unmanned aerial vehicle-based remote sensing), has attracted much attention because it offers improved accuracy gains compared with genomic prediction based only on marker genotypes. Although there is a trade-off between accuracy gains and phenotyping costs of secondary traits, no attempt has been made to optimize these trade-offs. In this study, we propose a novel approach to optimize multivariate genomic prediction models for secondary traits measurable at early growth stages for improved accuracy gains and phenotyping costs. The proposed approach employs Bayesian optimization for efficient Pareto frontier estimation, representing the maximum accuracy at a given cost. The proposed approach successfully estimated the optimal secondary trait combinations across a range of costs while providing genomic predictions for only about 20% of all possible combinations. The simulation results reflecting the characteristics of each scenario of the simulated target traits showed that the obtained optimal combinations were reasonable. Analysis of real-time target trait data showed that the proposed multivariate genomic prediction model had significantly superior accuracy compared to the univariate genomic prediction model.

Background Risk prediction models for time-to-event outcomes play a vital role in personalized decision-making. A patient’s biomarker values, such as medical lab results, are often measured over time but traditional prediction models ignore their longitudinal nature, using only baseline information. Dynamic prediction incorporates longitudinal information to produce updated survival predictions during follow-up. Existing methods for dynamic prediction include joint modeling, which often suffers from computational complexity and poor performance under misspecification, and landmarking, which has a straightforward implementation but typically relies on a proportional hazards model. Random survival forests (RSF), a machine learning algorithm for time-to-event outcomes, can capture complex relationships between the predictors and survival without requiring prior specification and has been shown to have superior predictive performance. Methods We propose an alternative approach for dynamic prediction using random survival forests in a landmarking framework. With a simulation study, we compared the predictive performance of our proposed method with Cox landmarking and joint modeling in situations where the proportional hazards assumption does not hold and the longitudinal marker(s) have a complex relationship with the survival outcome. We illustrated the use of the RSF landmark approach in two clinical applications to assess the performance of various RSF model building decisions and to demonstrate its use in obtaining dynamic predictions. Results In simulation studies, RSF landmarking outperformed joint modeling and Cox landmarking when a complex relationship between the survival and longitudinal marker processes was present. It was also useful in application when there were several predictors for which the clinical relevance was unknown and multiple longitudinal biomarkers were present. Individualized dynamic predictions can be obtained from this method and the variable importance metric is useful for examining the changing predictive power of variables over time. In addition, RSF landmarking is easily implementable in standard software and using suggested specifications requires less computation time than joint modeling. Conclusions RSF landmarking is a nonparametric, machine learning alternative to current methods for obtaining dynamic predictions when there are complex or unknown relationships present. It requires little upfront decision-making and has comparable predictive performance and has preferable computational speed.

Background and purpose — Advancements in software and hardware have enabled the rise of clinical prediction models based on machine learning (ML) in orthopedic surgery. Given their growing popularity and their likely implementation in clinical practice we evaluated which outcomes these new models have focused on and what methodologies are being employed. Material and methods — We performed a systematic search in PubMed, Embase, and Cochrane Library for studies published up to June 18, 2020. Studies reporting on non-ML prediction models or non-orthopedic outcomes were excluded. After screening 7,138 studies, 59 studies reporting on 77 prediction models were included. We extracted data regarding outcome, study design, and reported performance metrics. Results — Of the 77 identified ML prediction models the most commonly reported outcome domain was medical management (17/77). Spinal surgery was the most commonly involved orthopedic subspecialty (28/77). The most frequently employed algorithm was neural networks (42/77). Median size of datasets was 5,507 (IQR 635–26,364). The median area under the curve (AUC) was 0.80 (IQR 0.73–0.86). Calibration was reported for 26 of the models and 14 provided decision-curve analysis. Interpretation — ML prediction models have been developed for a wide variety of topics in orthopedics. Topics regarding medical management were the most commonly studied. Heterogeneity between studies is based on study size, algorithm, and time-point of outcome. Calibration and decision-curve analysis were generally poorly reported.

Cryptocurrency is a new sort of asset that has emerged as a result of the advancement of financial technology and it has created a big opportunity for researches. Cryptocurrency price forecasting is difficult due to price volatility and dynamism. Around the world, there are hundreds of cryptocurrencies that are used. This paper proposes three types of recurrent neural network (RNN) algorithms used to predict the prices of three types of cryptocurrencies, namely Bitcoin (BTC), Litecoin (LTC), and Ethereum (ETH). The models show excellent predictions depending on the mean absolute percentage error (MAPE). Results obtained from these models show that the gated recurrent unit (GRU) performed better in prediction for all types of cryptocurrency than the long short-term memory (LSTM) and bidirectional LSTM (bi-LSTM) models. Therefore, it can be considered the best algorithm. GRU presents the most accurate prediction for LTC with MAPE percentages of 0.2454%, 0.8267%, and 0.2116% for BTC, ETH, and LTC, respectively. The bi-LSTM algorithm presents the lowest prediction result compared with the other two algorithms as the MAPE percentages are: 5.990%, 6.85%, and 2.332% for BTC, ETH, and LTC, respectively. Overall, the prediction models in this paper represent accurate results close to the actual prices of cryptocurrencies. The importance of having these models is that they can have significant economic ramifications by helping investors and traders to pinpoint cryptocurrency sales and purchasing. As a plan for future work, a recommendation is made to investigate other factors that might affect the prices of cryptocurrency market such as social media, tweets, and trading volume.

Objectives Vibrant Soundbridge (VSB) was developed for treatment of hearing loss, but clinical outcomes vary and prognostic factors predicting the success of the treatment remain unknown. We examined clinical outcomes of VSB for conductive or mixed hearing loss, prognostic factors by analyzing prediction models, and cut-off values to predict the outcomes. Study design Retrospective chart review. Setting Tertiary care hospital. Patients Thirty patients who underwent VSB surgery from January 2017 to December 2019 at our hospital. Intervention Audiological tests were performed prior to and 3 months after surgery; patients completed questionnaires 3 months after surgery. Main outcome measures We used a multiregression and the random forest algorithm for predictions. Mean absolute errors and coefficient of determinations were calculated to estimate prediction accuracies. Coefficient values in the multiregression model and the importance of features in the random forest model were calculated to clarify prognostic factors. Receiver operation characteristic curves were plotted. Results All audiological outcomes improved after surgery. The random forest model (mean absolute error: 0.06) recorded more accuracy than the multiregression model (mean absolute error: 0.12). Speech discrimination score in a silent context in patients with hearing aids was the most influential factor (coefficient value: 0.51, featured value: 0.71). The candidate cut-off value was 36% (sensitivity: 89%, specificity: 75%). Conclusions VSB is an effective treatment for conductive or mixed hearing loss. Machine learning demonstrated more precise predictions, and speech discrimination scores in a silent context in patients with hearing aids were the most important factor in predicting clinical outcomes.

In the big data era, deep learning technologies have been widely exploited to mine increasingly available traffic data for mobile traffic prediction. Proactive management and optimization of mobile network resources for various wireless services require accurate mobile traffic prediction on different time and spacial scales. However, training deep learning models for different traffic prediction tasks individually is not only time consuming but also sometimes unrealistic as there are not always sufficient historical traffic records available. In this paper, we propose a novel mobile traffic prediction framework based on deep meta-learning (MTPFoDML), which can adaptively learn to learn the proper prediction model for each distinct prediction task from accumulated meta-knowledge of previous prediction tasks. In MTPFoDML, we regard each mobile traffic prediction task as a base-task and adopt a long short-term memory (LSTM) network with a fixed structure as the base-learner for each base-task. By transforming real-world mobile traffic data into the frequency domain, we find that the five main frequency components can characterize the mobile traffic variation over hours, days, and weeks, hence can be used as meta-features of a base-task. We employ a multi-layer perceptron (MLP) as the meta-learner to find the optimal super-parameter value and initial training status for the base-learner of each new base-task according to its meta-features, thus improving the base-learner's prediction accuracy and learning efficiency. Extensive experiments using real-world mobile traffic datasets demonstrate that our framework outperforms the existing prediction models for the same size of base-task training sets. Moreover, while guaranteeing a similar or even better prediction accuracy, meta-learning in MTPFoDML can lead to about 75% and 81%reduction in epoches and base-samples needed to train the base-learners, respectively.