Search

Multi-view classification with limited sample size and data augmentation is a very common machine learning (ML) problem in medicine. With limited data, a triplet network approach for two-stage representation learning has been proposed. However, effective training and verifying the features from the representation network for their suitability in subsequent classifiers are still unsolved problems. Although typical distance-based metrics for the training capture the overall class separability of the features, the performance according to these metrics does not always lead to an optimal classification. Consequently, an exhaustive tuning with all feature-classifier combinations is required to search for the best end result. To overcome this challenge, we developed a novel nearest-neighbor (NN) validation strategy based on the triplet metric. This strategy is supported by a theoretical foundation to provide the best selection of the features with a lower bound of the highest end performance. The proposed strategy is a transparent approach to identify whether to improve the features or the classifier. This avoids the need for repeated tuning. Our evaluations on real-world medical imaging tasks (i.e., radiation therapy delivery error prediction and sarcoma survival prediction) show that our strategy is superior to other common deep representation learning baselines [i.e., autoencoder (AE) and softmax]. The strategy addresses the issue of feature's interpretability which enables more holistic feature creation such that the medical experts can focus on specifying relevant data as opposed to tedious feature engineering.

With the ubiquitous deployment and applications of internet of things (IoT), security issues pose a critical challenge to IoT devices. External attackers often utilize vulnerable IoT devices to invade the target’s internal network and then further cause a security threat to the whole network. To prevent such attacks, it is necessary to develop a security mechanism to control the access of suspicious IoT devices and manage the internal devices. In recent years, deep learning (DL) algorithm has been widely used in the field of edge device identification (EDI), and has made great achievements. However, these previous methods are essentially centralized learning-based EDI (CentEDI) that trains all data together, which can not guarantee data security and not conducive to deployment on edge devices. To address this problem, we introduce a federated learning-based EDI (FedeEDI) method via network traffic to automatically identify edge devices connected to the whole network. Experimental results show that the training efficiency of our proposed FedeEDI method is much higher than that of the CentEDI method, although its classification accuracy is slightly reduced. In contrast to the CentEDI method, the proposed FedeEDI method has two main advantages: faster training speed and safer training process.

Android overlay enables one app to draw over other apps by creating an extra View layer atop the host View, which nevertheless can be exploited by malicious apps (malware) to attack users. To combat this threat, prior countermeasures concentrate on restricting the capabilities of overlays at the OS level while sacrificing overlays usability; recently, the overlay mechanism has been substantially updated to prevent a variety of attacks, which however can still be evaded by considerable adversaries. To address these shortcomings, a more pragmatic approach is to enable early detection of overlay-based malware during the app market review process, so that all the capabilities of overlays can stay unchanged. For this purpose, in this paper we first conduct a large-scale comparative study of overlay characteristics in benign and malicious apps, and then implement the OverlayChecker system to automatically detect overlay-based malware for one of the worlds largest Android app stores. In particular, we have made systematic efforts in feature engineering, UI exploration, emulation architecture, and run-time environment, thus maintaining high detection accuracy (97% precision and 97% recall) and short per-app scan time (1.7 minutes) with only two commodity servers, under an intensive workload of 10K newly submitted apps per day.

In recent years, Botnet has become one of the most dreadful type of malicious entity. Because of the hidden and carrying capacity of Botnet, the detection task has become a real challenge. Different methodologies have been applied for finding the source of Botnet at an early stage. Machine Learning and Deep Learning have greatly impacted these Botnet detection methodologies. But still, it is a difficult task because of the lesser number of features available in the Botnet datasets. In this paper, we have proposed a Fuzzy Logic based feature engineering method. The proposed method first identifies the fuzzy elements in the dataset and then generates fuzzy sets. The features generated using this method is used by an Artificial Neural Network for classification of Botnet. To train and evaluate the ANN model, we have used the CTU-13 dataset. The proposed feature engineering method and Botnet classification method has performed well with an accuracy of 99.94%. Still this method needs to be tested on different datasets. In future, new fuzzy rules can also be made to generate a new set of features as well as these rules can be used to generate features from other datasets.

One truism of deep learning is that the automatic feature engineering (seen in the first layers of those networks) excuses data scientists from performing tedious manual feature engineering prior to running DL. For the specific case of deep learning for defect prediction, we show that that truism is false. Specifically, when we preprocess data with a novel oversampling technique called fuzzy sampling, as part of a larger pipeline called GHOST (Goal-oriented Hyper-parameter Optimization for Scalable Training), then we can do significantly better than the prior DL state of the art in 14/20 defect data sets. Our approach yields state-of-the-art results significantly faster deep learners. These results present a cogent case for the use of oversampling prior to applying deep learning on software defect prediction datasets., Comment: v3, revision 2 (minor revision); submitted to TSE

Bike-sharing systems have made notable contributions to cities by providing green and sustainable mobility service to users. Over the years, many studies have been conducted to understand or anticipate the usage of these systems, with the hope to inform their future developments. One important task is to accurately predict usage patterns of the systems. Although many deep learning algorithms have been developed in recent years to support travel demand forecast, they have mainly been used to predict traffic volume or speed on roadways. Few studies have applied them to bike-sharing systems. Moreover, these studies usually focus on one single dataset or study area. The effectiveness and robustness of the prediction algorithms are not systematically evaluated. In this study, we propose a Spatial-Temporal Memory Network (STMN) to predict short-term usage of bicycles in bike-sharing systems. The framework employs Convolutional Long Short-Term Memory models and a feature engineering technique to capture the spatial-temporal dependencies in historical data for the prediction task. Four testing sites are used to evaluate the model. These four sites include two station-based systems (Chicago and New York) and two dockless bike-sharing systems (Singapore and New Taipei City). By assessing STMN with several baseline models, we find that STMN achieves the best overall performance in all the four cities. The model also achieves superior performance in urban areas with varying levels of bicycle usage and during peak periods when demand is high. The findings suggest the reliability of STMN in predicting bicycle usage for different types of bike-sharing systems.

Stereo matching is a challenging research topic in driving assistance systems. Existing stereo matching methods work well under normal day-light conditions. However, they fail to operate under adverse driving conditions, such as at night and during snowfall. This paper proposes a robust stereo matching framework using both deep-learning-based features and feature engineering. The proposed method investigates the benefits of features based on feature engineering and deep learning for solving stereo matching problems. Robust feature engineering is proposed for handling specific driving under adverse weather conditions, and a robust feature based on deep learning is considered for handling unspecific driving under extreme weather conditions. The proposed study has shown significantly improved accuracy by 8.31% for the state-of-the-art census based on semi-global matching under the reflection regions using the KITTI Stereo 2012 benchmark. Moreover, the experimental results demonstrate that the proposed system obtains more stable results than existing stereo methods based on deep learning on various stereo datasets, such as the Middlebury, EISAT, HCI, and CCD datasets.

Most of the classification methods of urban functional areas nowadays are only based on single source data analysis and modeling, which can not make full use of the multi-scale and multi-source data that is easy to obtain. Therefore, this paper proposed a classification model of urban functional areas based on multi-modal machine learning, by analyzing regional remote sensing images and behavior data of visitors in the area, using the combination of supervised methods extracted the deep-seated features and relationships of kinds of data, filtered and merged the overall and local features of the data. The model used dual branch neural network combining SE-ResNeXt and Dual Path Network (DPN) to automatically mined and fused the overall characteristics of multi-source data, and used the designed feature engineering to deeply mine the behavior data of users to obtain more association information, then combined the algorithm based on Gradient Boosting Decision Tree to learn the characteristics of different levels and obtained the classification probability for different levels of features. Finally, we continued to use the algorithm based on the Gradient Boosting Decision Tree to learn the probability distribution of different levels of features to obtain the final prediction results of urban functional area classification. Through the analysis and experimental verification of real data sets, the results showed that MM-UrbanFAC model can effectively integrate the features of multi-modal data. Compared with a single classifier, the integration framework based on gradient lifting tree improved the prediction performance, this method can effectively integrate the results of multiple models and accurately classify urban functional areas, and the model can provide reference for tourism recommendation, urban land planning and urban construction.

Improvement in individual mechanical properties of carbon steels, such as strength or ductility, can no longer keep up with the increasingly demanding service environment. Therefore, it is of practical significance to improve two or more mechanical properties accurately and efficiently. In this work, five machine learning algorithms are first employed to establish prediction models for different mechanical properties (tensile strength, fracture strength, Charpy absorbed energy, hardness, fatigue strength, and elongation) based on the collected carbon steels data. Then, a set of mutually exclusive properties (tensile strength and elongation) and the key descriptors of the corresponding properties are identified by feature engineering, and the importance of the key materials descriptors is analyzed. The prediction models based on key descriptors for tensile strength and elongation also demonstrate good accuracy. All the key descriptors are considered as input features for the comprehensive performance (CP) calculated from the product of tensile strength and elongation. Finally, we develop a machine learning prediction model for CP and successfully apply the efficient global optimization algorithm to optimize two mutually exclusive mechanical properties. This work provides a new multi-objective optimization strategy that is expected to be used for the development of new steels with excellent comprehensive performance.

Computational drug design relies on the calculation of binding strength between two biological counterparts especially a chemical compound, i.e., a ligand, and a protein. Predicting the affinity of protein-ligand binding with reasonable accuracy is crucial for drug discovery, and enables the optimization of compounds to achieve better interaction with their target protein. In this paper, we propose a data-driven framework named DeepAtom to accurately predict the protein-ligand binding affinity. With 3D Convolutional Neural Network (3D-CNN) architecture, DeepAtom could automatically extract binding related atomic interaction patterns from the voxelized complex structure. Compared with the other CNN based approaches, our light-weight model design effectively improves the model representational capacity, even with the limited available training data. We carried out validation experiments on the PDBbind v.2016 benchmark and the independent Astex Diverse Set. We demonstrate that the less feature engineering dependent DeepAtom approach consistently outperforms the other baseline scoring methods. We also compile and propose a new benchmark dataset to further improve the model performances. With the new dataset as training input, DeepAtom achieves Pearson's R=0.83 and RMSE=1.23 pK units on the PDBbind v.2016 core set. The promising results demonstrate that DeepAtom models can be potentially adopted in computational drug development protocols such as molecular docking and virtual screening.

Content delivery networks (CDNs) provide efficient content distribution over the Internet. CDNs improve the connectivity and efficiency of global communications, but their caching mechanisms may be breached by cyber-attackers. Among the security mechanisms, effective anomaly detection forms an important part of CDN security enhancement. In this work, we propose a multi-perspective unsupervised learning framework for anomaly detection in CDNs. In the proposed framework, a multi-perspective feature engineering approach, an optimized unsupervised anomaly detection model that utilizes an isolation forest and a Gaussian mixture model, and a multi-perspective validation method, are developed to detect abnormal behaviors in CDNs mainly from the client Internet Protocol (IP) and node perspectives, therefore to identify the denial of service (DoS) and cache pollution attack (CPA) patterns. Experimental results are presented based on the analytics of eight days of real-world CDN log data provided by a major CDN operator. Through experiments, the abnormal contents, compromised nodes, malicious IPs, as well as their corresponding attack types, are identified effectively by the proposed framework and validated by multiple cybersecurity experts. This shows the effectiveness of the proposed method when applied to real-world CDN data., Comment: Accepted and to Appear in IEEE Transactions on Network and Service Management

Quantitative ultrasound (QUS), which is commonly used to extract quantitative features from the ultrasound radiofrequency (RF) data or the RF envelope signals for tissue characterization, is becoming a promising technique for noninvasive assessments of liver fibrosis. However, the number of feature variables examined and finally used in the existing QUS methods is typically small, to some extent limiting the diagnostic performance. Therefore, this paper devises a new multiparametric QUS (MP-QUS) method which enables the extraction of a large number of feature variables from US RF signals and allows for the use of feature-engineering and machinelearning based algorithms for liver fibrosis assessment. In the MP-QUS, eighty-four feature variables were extracted from multiple QUS parametric maps derived from the RF signals and the envelope data. Afterwards, feature reduction and selection were performed in turn to remove the feature redundancy and identify the best combination of features in the reduced feature set. Finally, a variety of machine-learning algorithms were tested for classifying liver fibrosis with the selected features, based on the results of which the optimal classifier was established and used for final classification. The performance of the proposed MPQUS method for staging liver fibrosis was evaluated on an animal model, with histologic examination as the reference standard. The mean accuracy, sensitivity, specificity and area under the receiver-operating-characteristic curve achieved by MP-QUS are respectively 83.38%, 86.04%, 80.82% and 0.891 for recognizing significant liver fibrosis, and 85.50%, 88.92%, 85.24% and 0.924 for diagnosing liver cirrhosis. The proposed MP-QUS method paves a way for its future extension to assess liver fibrosis in human subjects.

Text segmentation is a fundamental task in natural language processing. Depending on the levels of granularity, the task can be defined as segmenting a document into topical segments, or segmenting a sentence into elementary discourse units (EDUs). Traditional solutions to the two tasks heavily rely on carefully designed features. The recently proposed neural models do not need manual feature engineering, but they either suffer from sparse boundary tags or cannot efficiently handle the issue of variable size output vocabulary. In light of such limitations, we propose a generic end-to-end segmentation model, namely SEGBOT, which first uses a bidirectional recurrent neural network to encode an input text sequence. SEGBOT then uses another recurrent neural networks, together with a pointer network, to select text boundaries in the input sequence. In this way, SEGBOT does not require any hand-crafted features. More importantly, SEGBOT inherently handles the issue of variable size output vocabulary and the issue of sparse boundary tags. In our experiments, SEGBOT outperforms state-of-the-art models on two tasks: document-level topic segmentation and sentence-level EDU segmentation. As a downstream application, we further propose a hierarchical attention model for sentence-level sentiment analysis based on the outcomes of SEGBOT. The hierarchical model can make full use of both word-level and EDU-level information simultaneously for sentence-level sentiment analysis. In particular, it can effectively exploit EDU-level information, such as the inner properties of EDUs, which cannot be fully encoded in word-level features. Experimental results show that our hierarchical model achieves new state-of-the-art results on the Movie Review and Stanford Sentiment Treebank benchmarks.

In modern neuroscience and clinical study, neuroscientists and clinicians often use non-invasive imaging techniques to validate theories and computational models, observe brain activities and diagnose brain disorders. The functional Magnetic Resonance Imaging (fMRI) is one of the commonly-used imaging modalities that can be used to understand human brain mechanisms as well as the diagnosis and treatment of brain disorders. The advances in artificial intelligence and the emergence of deep learning techniques have shown promising results to better interpret fMRI data. Deep learning techniques have rapidly become the state of the art for analyzing fMRI data sets and resulted in performance improvements in diverse fMRI applications. Deep learning is normally presented as an end-to-end learning process and can alleviate feature engineering requirements and hence reduce domain knowledge requirements to some extent. Under the framework of deep learning, fMRI data can be considered as images, time series or images series. Hence, different deep learning models such as convolutional neural networks, recurrent neural network, or a combination of both, can be developed to process fMRI data for different tasks. In this review, we discussed the basics of deep learning methods and focused on its successful implementations for brain disorder diagnosis based on fMRI images. The goal is to provide a high-level overview of brain disorder diagnosis with fMRI images from the perspective of deep learning applications.

Predictive maintenance (PdM) advocates for the usage of machine learning technologies to monitor asset's health conditions and plan maintenance activities accordingly. However, according to the specific degradation process, some health-related measures (e.g. temperature) may be not informative enough to reliably assess the health stage. Moreover, each measure needs to be properly treated to extract the information linked to the health stage. Those issues are usually addressed by performing a manual feature engineering, which results in high management cost and poor generalization capability of those approaches. In this work, we address this issue by coupling a health stage classifier with a feature learning mechanism. With feature learning, minimally processed data are automatically transformed into informative features. Many effective feature learning approaches are based on deep learning. With those, the features are obtained as a non-linear combination of the inputs, thus it is difficult to understand the input's contribution to the classification outcome and so the reasoning behind the model. Still, these insights are increasingly required to interpret the results and assess the reliability of the model. In this regard, we propose a feature learning approach able to (i) effectively extract high-quality features by processing different input signals, and (ii) provide useful insights about the most informative domain transformations (e.g. Fourier transform or probability density function) of the input signals (e.g. vibration or temperature). The effectiveness of the proposed approach is tested with publicly available real-world datasets about bearings' progressive deterioration and compared with the traditional feature engineering approach.

Feature engineering is one of the major challenges of machine learning . While multiple automation solutions have been proposed in recent years, the vast majority focuses on extracting features from the analyzed dataset itself and not from other (external) sources. In this study we present FGSES, a general framework for automatic feature engineering and its application to DBpedia. Our framework automatically matches the entities in the analyzed dataset to those of the external data source, and then proceeds to generate a large and diverse set of candidate features, both from structured and unstructured content. To efficiently process the large number of generated features, FGSES uses a meta learning-based ranking approach. Our evaluation, conducted on 18 tabular datasets with diverse characteristics, shows that FGSES achieves an average error reduction of 16.5%, significantly outperforming the evaluated baselines.

Existing decision support system frameworks for diagnosing Parkinson's disease (PD) through handwriting, speech, or gait characteristics share very similar pipelines. Although in some cases, patient data can be captured by commercially available devices, specialized devices or even custom-made prototypes are often required for such tasks. Captured data are used for extracting features that are carefully designed on the basis of domain and problem knowledge. These features are then fed to classifiers that provide a final decision. In this article, we present an approach in which end-to-end processing by a convolutional neural network (CNN) is utilized to diagnose PD from handwriting images, without the use of additional signals. This eliminates any need for specialized devices or feature engineering. To improve the performance of the proposed pretrained CNN, we propose the idea of multiple fine tuning to bridge the gap between semantically different source and target datasets and facilitate more efficient transfer learning. The proposed architecture, which is based on multiple fine tuning and an ensemble of multiple-fine-tuned CNNs, achieves 94.7% accuracy in the classification of PD from offline handwriting.

The Internet of Things (IoT) is becoming an indispensable part of everyday life, enabling a variety of emerging services and applications. However, the presence of rogue IoT devices has exposed the IoT to untold risks with severe consequences. The first step in securing the IoT is detecting rogue IoT devices and identifying legitimate ones. Conventional approaches use cryptographic mechanisms to authenticate and verify legitimate devices’ identities. However, cryptographic protocols are not available in many systems. Meanwhile, these methods are less effective when legitimate devices can be exploited or encryption keys are disclosed. Therefore, non-cryptographic IoT device identification and rogue device detection become efficient solutions to secure existing systems and will provide additional protection to systems with cryptographic protocols. Non-cryptographic approaches require more effort and are not yet adequately investigated. In this paper, we provide a comprehensive survey on machine learning technologies for the identification of IoT devices along with the detection of compromised or falsified ones from the viewpoint of passive surveillance agents or network operators. We classify the IoT device identification and detection into four categories: device-specific pattern recognition, Deep Learning enabled device identification, unsupervised device identification, and abnormal device detection. Meanwhile, we discuss various ML-related enabling technologies for this purpose. These enabling technologies include learning algorithms, feature engineering on network traffic traces and wireless signals, incremental learning, and abnormality detection.

Malicious software (malware) is a major threat to the systems and networks’ security. Although anti-malware products are used to protect systems and networks against malware attacks, obfuscated malware that is capable of evading analysis and detection by anti-malware software have become prevalent. Therefore, how to detect and remove obfuscated malware from the systems has become a major concern. In this research work, we propose a semi-supervised approach that integrates deep learning, feature engineering, image transformation and processing techniques for obfuscated malware detection. We validated the proposed approach through experiments and compared it with existing approaches. With 99.12% accuracy in detecting obfuscated malware detection, the proposed approach substantially outperformed the other approaches.

Offshore wind is a rapidly maturing renewable energy that has presented a large growth over the last decade. This increase in offshore wind capacity has led to the need for more effective monitoring strategies, as currently, Operation and Maintenance (O&M) costs make up to 30% of the overall cost of energy. This study presented a novel data-driven approach to condition monitoring systems by utilizing the existing Supervisory Control And Data Acquisition (SCADA) system and integrating a wide range of machine learning and data mining techniques namely: data pre-processing & re-sampling, anomalies detection & treatment, feature engineering, and hyperparameter optimization, to design a Normal Behaviour Model of the generator for fault detection purposes. An ensemble model of the Extreme Gradient Boosting (XGBoost) framework was successfully developed and critically compared with a Long Short-Term Memory (LSTM) deep learning neural network. The results showed that, in terms of temperature prediction, the proposed methodology captures a high level of accuracy at low computational costs. Moreover, it can be concluded that XGBoost outperformed LSTM in predictive accuracy whilst requiring smaller training times and showcasing a smaller sensitivity to noise that existed in the SCADA database.

With ever-increasing, internet-connected devices provide an opportunity to fulfil the attacker’s malicious intention. They use malicious programs to compromise the devices and use them to infect others also. The security researchers are straggling to develop a technique that detects all the malware accurately because of the use of invincible techniques in the development of malware such as strong encryption, obfuscation, polymorphic and metamorphic engine. In this context, this paper proposes a novel malware classification with fine-tune convolution neural networks (MCFT-CNN) model. The MCFT-CNN model detects the unknown malware sample without feature engineering and prior knowledge of binary code analysis or reverse engineering, even the advanced evading techniques used to develop the malware. The model uses deep transfer learning to classify the malware images to their respective malware families. The proposed model enhances the ResNet50 model by altering the last layer with a fully connected dense layer. The output of fully connected dense layer and knowledge of ImageNet model are supplied to softmax layer for malware classification. The model is trained with MalImg malware datasets. The proposed model reported 99.18% accuracy and 5.14ms prediction time. The model also shows consistent performance with a relatively larger dataset (Microsoft malware challenge dataset, approximately 500GB) with 98.63% accuracy and 5.15ms prediction time. The proposed model shows consistent efficacy with two benchmark datasets that clarify the model’s generalisability to perform on the diverse datasets.

Distributed Denial of Service (DDoS) is a predominant threat to the availability of online services due to their size and frequency. However, developing an effective security mechanism to protect a network from this threat is a big challenge because DDoS uses various attack approaches coupled with several possible combinations. Furthermore, most of the existing deep learning- (DL-) based models pose a high processing overhead or may not perform well to detect the recently reported DDoS attacks as these models use outdated datasets for training and evaluation. To address the issues mentioned earlier, we propose CyDDoS, an integrated intrusion detection system (IDS) framework, which combines an ensemble of feature engineering algorithms with the deep neural network. The ensemble feature selection is based on five machine learning classifiers used to identify and extract the most relevant features used by the predictive model. This approach improves the model performance by processing only a subset of relevant features while reducing the computation requirement. We evaluate the model performance based on CICDDoS2019, a modern and realistic dataset consisting of normal and DDoS attack traffic. The evaluation considers different validation metrics such as accuracy, precision, F1-Score, and recall to argue the effectiveness of the proposed framework against state-of-the-art IDSs.

Text classification is a fundamental language task in Natural Language Processing. A variety of sequential models are capable of making good predictions, yet there is a lack of connection between language semantics and prediction results. This paper proposes a novel influence score (I-score), a greedy search algorithm, called Backward Dropping Algorithm (BDA), and a novel feature engineering technique called the “dagger technique”. First, the paper proposes to use the novel influence score (I-score) to detect and search for the important language semantics in text documents that are useful for making good predictions in text classification tasks. Next, a greedy search algorithm, called the Backward Dropping Algorithm, is proposed to handle long-term dependencies in the dataset. Moreover, the paper proposes a novel engineering technique called the “dagger technique” that fully preserves the relationship between the explanatory variable and the response variable. The proposed techniques can be further generalized into any feed-forward Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs), and any neural network. A real-world application on the Internet Movie Database (IMDB) is used and the proposed methods are applied to improve prediction performance with an 81% error reduction compared to other popular peers if I-score and “dagger technique” are not implemented.

Machine learning (ML) is emerging as a promising solution for managing the physical layer complexity of heterogeneous dynamic optical networks transporting multiple applications in a software defined network (SDN) context, namely for performance prediction. We propose two multivariate neural network models based on gated recurrent unit (GRU) and long short-term memory (LSTM) methods, trained with field performance data and features, for predicting lightpath signal-to-noise ratio (SNR) over forecast horizons of up to 4 days. The best performance is achieved by using a 5-feature LSTM multivariate model over forecast horizons of up to 96 hours, with an absolute maximum error (AME) of 0.90 dB, compared to 0.91 dB and 0.97 dB for the GRU and LSTM univariate models, respectively, and 1.21 dB for a persistence model. The 2-feature multivariate models obtained through feature engineering perform better than their univariate counterparts for forecast horizons of up to 40 hours. Lastly, we explore the concept of transfer learning (TL) by testing the trained multivariate LSTM and univariate GRU models on field data from two lightpaths carried on the same route. The TL models underperform the naive model for the lightpath carried in a different optical fiber. However, for the lightpath carried in the same optical fiber on a portion of the same route, the LSTM-based TL model outperforms the naive model with a difference of up to 0.11 dB at a 96-hour forecast horizon, compared to 0.30 dB for the lightpath in the source domain, while using 3 times less training data.

Click-through rate (CTR) prediction plays an important role in many industrial applications, feature engineering directly influences CTR prediction performance because features are normally the multi-field type. However, the existing CTR prediction techniques either neglect the importance of each feature or regard the feature interactions equally for feature learning. In addition, using an inner product or a Hadamard product is too simple to effectively model the feature interactions. These limitations lead to suboptimal performances of existing models. In this paper, we propose a framework called Hierarchical Attention and Feature Projection neural network (HAFP) for CTR prediction, which enables the automatically learning of more representative and efficient feature representation in an end-to-end manner. Towards this end, we employ a feature learning layer with a hierarchical attention mechanism to jointly extract more generalized and dominant features and feature interactions. In addition, a projective bilinear function is designed in meaningful second-order interaction encoder to effectively learn more fine-grained and comprehensive second-order feature interactions. Taking advantages of the hierarchical attention mechanism and the projective bilinear function, our proposed model can not only model feature learning in a flexible fashion, but also provide an interpretable capability of the prediction results. Experimental results on two real-world datasets demonstrate that HAFP outperforms the state-of-the-art in terms of Logloss and AUC for CTR prediction baselines. Further analysis verifies the importance of the proposed hierarchical attention mechanism and the projective bilinear function for modelling the feature representation, showing the rationality and effectiveness of HAFP.

Machine learning is an important tool for extracting information from medical images. Deep learning has made this more efficient by not requiring an explicit feature extraction step and in some cases detecting features that humans had not identified. The rapid advance of deep learning technologies continues to result in valuable tools. The most effective use of these tools will occur when developers also understand the properties of medical images and the clinical questions at hand. The performance metrics also are critical for guiding the training of an artificial intelligence and for assessing and comparing its tools.

This paper presents a novel supervised machine learning-based electric theft detection approach using the feature engineered-CatBoost algorithm in conjunction with the SMOTETomek algorithm. Contrary to the previous literature, where the missing observations in data are either ignored or imputed with average values, this work utilizes k-Nearest neighbor technique for missing data imputation; thus, an accurate and realistic estimation of the missing data is achieved. To mitigate the biasness to the majority data class, the proposed model utilizes the SMOTETomek algorithm, which neutralizes the mentioned effect by managing a proper balance between over-sampling and under-sampling techniques. Feature Extraction and Scalable Hypothesis (FRESH) algorithm is utilized at the later stage of the proposed NTL detection framework to extract and select the most relevant data features from the provided dataset. Afterward, the model is trained using the CatBoost algorithm to classify the consumers into two distinct categories, i.e., genuine and theft. Finally, to interpret the model’s decision for the corresponding predictions, the tree-SHAP algorithm is utilized. To validate the efficacy of the proposed ML based theft detection approach, its performance is compared with that of the traditional gradient boosting ML algorithms such as XGBoost, lightGBM, Ensemble bagging, boosting ML models, and other conventional ML models using five of the most widely used performance metrics, i.e., precision, accuracy, F1score Kappa and MCC. The proposed technique achieved an accuracy of 93% and a detection rate of 92%, which is significantly higher than all the considered competing algorithms under identical dataset and hyperparameters.

Automatic seizure onset zone (SOZ) localization using interictal electrocorticogram (ECoG) improves the diagnosis and treatment of patients with medically refractory epilepsy. This study aimed to investigate the characteristics of phase-amplitude coupling (PAC) extracted from interictal ECoG and the feasibility of PAC serving as a promising biomarker for SOZ identification. We employed the mean vector length modulation index approach on the 20-s ECoG window to calculate PAC features between low-frequency rhythms (0.5–24 Hz) and high frequency oscillations (HFOs) (80–560 Hz). We used statistical measures to test the significant difference in PAC between the SOZ and non-seizure onset zone (NSOZ). To overcome the drawback of handcraft feature engineering, we established novel machine learning models to learn automatically the characteristics of the obtained PAC features and classify them to identify the SOZ. Besides, to handle imbalanced dataset classification, we introduced novel feature-wise/class-wise re-weighting strategies in conjunction with classifiers. In addition, we proposed a time-series nest cross-validation to provide more accurate and unbiased evaluations for this model. Seven patients with focal cortical dysplasia were included in this study. The experiment results not only showed that a significant coupling at band pairs of slow waves and HFOs exists in the SOZ when compared with the NSOZ, but also indicated the effectiveness of the PAC features and the proposed models in achieving better classification performance .

© The Author(s) 2021.Emotion classification is a research field that aims to detect the emotions in a text using machine learning methods. In traditional machine learning (TML) methods, feature engineering processes cause the loss of some meaningful information, and classification performance is negatively affected. In addition, the success of modelling using deep learning (DL) approaches depends on the sample size. More samples are needed for Turkish due to the unique characteristics of the language. However, emotion classification data sets in Turkish are quite limited. In this study, the pretrained language model approach was used to create a stronger emotion classification model for Turkish. Well-known pretrained language models were fine-tuned for this purpose. The performances of these fine-tuned models for Turkish emotion classification were comprehensively compared with the performances of TML and DL methods in experimental studies. The proposed approach provides state-of-the-art performance for Turkish emotion classification.

In the field of polymer informatics, utilizing machine learning (ML) techniques to evaluate the glass transition temperature Tg and other properties of polymers has attracted extensive attention. This data-centric approach is much more efficient and practical than the laborious experimental measurements when encountered a daunting number of polymer structures. Various ML models are demonstrated to perform well for Tg prediction. Nevertheless, they are trained on different data sets, using different structure representations, and based on different feature engineering methods. Thus, the critical question arises on selecting a proper ML model to better handle the Tg prediction with generalization ability. To provide a fair comparison of different ML techniques and examine the key factors that affect the model performance, we carry out a systematic benchmark study by compiling 79 different ML models and training them on a large and diverse data set. The three major components in setting up an ML model are structure representations, feature representations, and ML algorithms. In terms of polymer structure representation, we consider the polymer monomer, repeat unit, and oligomer with longer chain structure. Based on that feature, representation is calculated, including Morgan fingerprinting with or without substructure frequency, RDKit descriptors, molecular embedding, molecular graph, etc. Afterward, the obtained feature input is trained using different ML algorithms, such as deep neural networks, convolutional neural networks, random forest, support vector machine, LASSO regression, and Gaussian process regression. We evaluate the performance of these ML models using a holdout test set and an extra unlabeled data set from high-throughput molecular dynamics simulation. The ML model's generalization ability on an unlabeled data set is especially focused, and the model's sensitivity to topology and the molecular weight of polymers is also taken into consideration. This benchmark study provides not only a guideline for the Tg prediction task but also a useful reference for other polymer informatics tasks.

While the open-source software development model has led to successful large-scale collaborations in building software systems, data science projects are frequently developed by individuals or small teams. We describe challenges to scaling data science collaborations and present a conceptual framework and ML programming model to address them. We instantiate these ideas in Ballet, the first lightweight framework for collaborative, open-source data science through a focus on feature engineering, and an accompanying cloud-based development environment. Using our framework, collaborators incrementally propose feature definitions to a repository which are each subjected to software and ML performance validation and can be automatically merged into an executable feature engineering pipeline. We leverage Ballet to conduct a case study analysis of an income prediction problem with 27 collaborators, and discuss implications for future designers of collaborative projects.

Real‐world Artificial Intelligence (AI) solutions need different tools for managing data, deployed at various scales. AI enables the machine to understand, make interactions, and communicate with humans in a human way. Machine learning (ML) can be termed as an application of AI that offers systems and the ability to learn and improve automatically from the experiences without explicit programming. In deep learning (DL), the computer will take inputs or learn from sound, images, or text and perform the action accordingly. There can be many examples of DL applications that are currently getting known: aerospace and defense, driving cars, medical, industrial automation, and electronics. The comparison between ML and DL can be done on the following basis: data dependencies, hardware dependencies, problem‐solving approach, feature engineering, execution time, and interpretability. Virtual assistants based on ML help people on many platforms, such as: mobile applications, smartphones taking multiple commands, and smart speakers.

Research in AI has proved to be revolutionarily beneficial to humankind from the past few decades. Many supporting techniques have been developed that indirectly evolved AI and directly enhanced various machine learning models; one of them being the feature engineering. It can also be considered as applied-ML. Another is the so-called feature selection which is method in which most contributing feature to final decision making, out of the entire feature space are selected for processing into an ML model. It is not an easy task to precisely calculate the dependency of the output variable onto the candidate features, particularly when the data is high in dimensions. In this regard, this study proposes a novel method named the cisoidal analysis based feature selection (CAFS) which uses both pre-established algorithms as well as a new approach of relating members of feature space to a complex sinusoid (cisoid) mathematically, then using signal processing techniques to eliminate certain elements in the entire feature space for enhanced feature selection and hence to obtain higher classification accuracy. Derived from experiments with five high dimensional datasets, CAFS displays significantly competitive performance than some of the pre-existing algorithms. CAFS is highly advantageous in reducing dimension of feature space in most of the applications.

Named entity recognition (NER) is crucial in various natural language processing (NLP) tasks. However, the nested entities which are common in practical corpus are often ignored in most of current NER models. To extract the nested entities, two categories of models (i.e., feature-based and neural network-based approaches) are proposed. However, the feature-based models suffer from the complicated feature engineering and often heavily rely on the external resources. Discarding the heavy feature engineering, recent neural network-based methods which treat the nested NER as a classification task are designed but still suffer from the heavy class imbalance issue and the high computational cost. To solve these problems, we propose a neural multi-task model with two modules: Binary Sequence Labeling and Candidate Region Classification to extract the nested entities. Extensive experiments are conducted on the public datasets. Comparing with recent neural network-based approaches, our proposed model achieves the better performance and obtains the higher efficiency.

Predicting materials properties from composition or structure is of great interest to the materials science community. Deep learning has recently garnered considerable interest in materials predictive tasks with low model errors when dealing with large materials data. However, deep learning models suffer in the small data regime that is common in materials science. Here we leverage the transfer learning concept and the graph network deep learning framework and develop the AtomSets machine learning framework for consistent high model accuracy at both small and large materials data. The AtomSets models can work with both compositional and structural materials data. By combining with transfer learned features from graph networks, they can achieve state-of-the-art accuracy from using small compositional data (130,000). The AtomSets models show much lower errors than the state-of-the-art graph network models at small data limits and the classical machine learning models at large data limits. They also transfer better in the simulated materials discovery process where the targeted materials have property values out of the training data limits. The models require minimal domain knowledge inputs and are free from feature engineering. The presented AtomSets model framework opens new routes for machine learning-assisted materials design and discovery.

Accurately recognizing various human actions is a key technique in many AI applications, such as visual tracking and human–computer interaction. Aiming at solving the difficulty that the local spatial–temporal features at low-layer is limited and the descriptiveness of middle-level features is weak, we propose a novel human action understanding framework by leveraging the spatial–temporal depth features. More specifically, based on the fact that violent action regions encode highly discriminative information during human action recognition, we employ human depth clue of video images to identify the salient regions for each human. The optical flow characteristics in the regions are utilized as the energy function to measure the regional representativeness. The deep learning architecture is proposed to identify the regions of motions by the energy function. In this way, the sample points are distributed in the areas with intense movements. The collected sample points are utilized as the learned features to capture human action. Based on the learned deep feature, an SVM classifier is used to identify different human action. Comprehensive experimental results have shown that the average recognition accuracy of our human action recognition algorithm reaches 92%, and also exhibits a high robustness to complicated backgrounds.

Application Programming Interfaces (APIs) have been widely discussed on social-technical platforms (e.g., Stack Overflow). Extracting API mentions from such informal software texts is the prerequisite for API-centric search and summarization of programming knowledge. Machine learning based API extraction has demonstrated superior performance than rule-based methods in informal software texts that lack consistent writing forms and annotations. However, machine learning based methods have a significant overhead in preparing training data and effective features. In this paper, we propose a multi-layer neural network based architecture for API extraction. Our architecture automatically learns character-, word- and sentence-level features from the input texts, thus removing the need for manual feature engineering and the dependence on advanced features (e.g., API gazetteers) beyond the input texts. We also propose to adopt transfer learning to adapt a source-library-trained model to a target-library, thus reducing the overhead of manual training-data labeling when the software text of multiple programming languages and libraries need to be processed. We conduct extensive experiments with six libraries of four programming languages which support diverse functionalities and have different API-naming and API-mention characteristics. Our experiments investigate the performance of our neural architecture for API extraction in informal software texts, the importance of different features, the effectiveness of transfer learning. Our results confirm not only the superior performance of our neural architecture than existing machine learning based methods for API extraction in informal software texts, but also the easy-to-deploy characteristic of our neural architecture.

In this work, we propose two deep learning-based architectures tailored for gas identification and quantification, which automatically tune hyper-parameters of the network for optimal performance. The immense success of deep learning in the field of computer vision and natural language processing inspired us to design deep learning-based gas identification and quantification network. The first architecture is proposed for gas quantification, which is based on 1D-CNN. It makes use of raw time-series gas sensor array data and provides the concentration of each gas in a mixture of gases. The second architecture is presented for gas quantification, which is based on a deep belief network combined with drift-aware feature adaptation strategy. The proposed models identify and quantify the gases with improved accuracy despite the presence of sensor drift. Additionally, hyper-parameters of both the networks are automatically tuned for optimal performance. Although several pattern recognition methods related to machine learning, fuzzy logic and hybrid models have been used to identify gas and quantify the gases in the mixture, the performances of these techniques enormously depend on the feature engineering and selection of hyper-parameters. Experimental results show that the proposed methods are an effective technique for identifying gases and quantifying the mixture of gases for e-nose data. We also present that the proposed methods outperforms various other methods and can provide higher identification and quantification accuracy in the pres-ence of sensor drift.

Blade icing detection becomes increasingly significant as it can avoid revenue loss and power degradation. Conventional methods are usually limited by additional costs, and model-driven methods heavily depend on prior domain knowledge. Data-driven methods, especially deep learning approaches without needing the time-consuming handcraft feature engineering, offer a promising solution for blade icing detection. However, the monitoring signals normally have complex and diverse features as wind turbine operates in complex environments, thus effective model is needed for data analyzing. Additionally, the distribution of monitoring data is imbalanced, which causes the abnormal data mining inadequate. In this work, a multilevel convolutional recurrent neural network (MCRNN), is proposed for blade icing detection. Specifically, discrete wavelet decomposition is leveraged to obtain multilevel features both from the time domain and the frequency domain. A parallel structure combining an LSTM branch and a CNN branch is established in each level for feature extraction. To alleviate the severe data imbalance, two mechanisms, including data resampling algorithm and class-rebalanced loss function, are investigated. Furthermore, a multi-step accumulation strategy is proposed to enhance the accuracy of real-time detection. Extensive studies demonstrate that the proposed MCRNN can achieve up 38.8% and 42.9% higher F1-score over the best baseline on the balanced data sets processed by data resampling algorithm and 23.9% and 30.6% higher on imbalanced data sets with MCRNN optimized by the class-rebalanced loss function. The real-time detection verifies the applicability of the proposed method and indicates that the proposed multi-step accumulation strategy can improve the accuracy of icing detection.

Automated machine learning (AutoML) has achieved remarkable progress on various tasks, which is attributed to its minimal involvement of manual feature and model designs. However, most of existing AutoML pipelines only touch parts of the full machine learning pipeline, e.g., neural architecture search or optimizer selection. This leaves potentially important components such as data cleaning and model ensemble out of the optimization, and still results in considerable human involvement and suboptimal performance. The main challenges lie in the huge search space assembling all possibilities over all components, as well as the generalization ability over different tasks like image, text, and tabular etc . In this paper, we present a first-of-its-kind fully AutoML pipeline, to comprehensively automate data preprocessing, feature engineering, model generation/selection/training and ensemble for an arbitrary dataset and evaluation metric. Our innovation lies in the comprehensive scope of a learning pipeline, with a novel “life-long” knowledge anchor design to fundamentally accelerate the search over the full search space. Such knowledge anchors record detailed information of pipelines and integrates them with an evolutionary algorithm for joint optimization across components. Experiments demonstrate that the result pipeline achieves state-of-the-art performance on multiple datasets and modalities. Specifically, the proposed framework was extensively evaluated in the NeurIPS 2019 AutoDL challenge, and won the only champion with a significant gap against other approaches, on all the image, video, speech, text and tabular tracks.

Next-generation sequencing techniques provide us with an opportunity for generating sequenced proteins and identifying the biological families and functions of these proteins. However, compared with identified proteins, uncharacterized proteins consist of a notable percentage of the overall proteins in the bioinformatics research field. Traditional family classification methods often devote themselves to extracting N-Gram features from sequences while ignoring motif information as well as affinity information between motifs and adjacent amino acids. Previous clustering-based algorithms have typically been used to define protein features with domain knowledge and annotate protein families based on extensive data samples. In this paper, we apply CNN based amino acid representation learning with limited characterized proteins to explore the performances of annotated protein families by taking into account the amino acid location information. Additionally, we apply the method to all reviewed protein sequences with their families retrieved from the UniProt database to evaluate our approach. Last but not least, we verify our model using those unreviewed protein records, which is typically ignored by other methods.

Personality Traits Detection is one of the important problems as a text analytics task in Natural Language Processing (NLP). Text analytics is the process of finding out insight knowledge over written text. Although most deep learning models give high performance, they often lack interpretability. Computer Vision (CV) has been affected significantly with inductive transfer learning, however training from scratch and task-specific modifications are still wanted in many NLP techniques. This paper addresses the problem of personality traits classification. We adopted the use of the Universal Language Model Fine-Tuning (ULMFiT) in personality traits detection. The model makes use of transfer learning rather than the classical shallow methods of word embedding and proved to be the most powerful model in many NLP problems. The basic advantage of using this model is that there is no need to do feature engineering before classification. When applied to benchmark dataset, the proposed method shows a statistical accuracy improvement of about 1% compared to the state-of-the-art results for the big five personality traits.

Nowadays radar systems have been facing with the disordered electromagnetic spectrum access and utilization in shared spectrum environments with radio communication systems. Numerous waveform recognition methods have been studied with feature engineering and conventional machine learning (ML) for intelligent radar systems, but they are critically challenged by practical problems of scalability and reliability. Deep learning (DL) with the ability to automatically learn the representational features is leveraged to handle the aforementioned obstacles effectively. In this work, we proposed a high-accurate waveform recognition method for intelligent radar systems by developing a novel residual-attention multiscale-accumulation convolutional network (RamNet). By deliberately incorporating the residual connection and attention connection in selective-feature improvement blocks, RamNet can enrich high-impact features without vanishing gradient. Moreover, a structural multiscale-accumulation connection is deployed to improve feature utilization by gathering the high-relevant features at multiple signal resolutions. Experimental results from exhaustive simulation demonstrate that RamNet recognizes waveform robustly under rigorous channel impairments and presents superior performance compared to traditional ML and state-of-the-art DL models.

In the future, vehicles will be equipped with increasingly advanced interactive intelligent electronic devices, which will induce drivers to conduct secondary tasks, thereby leading to distractions. Therefore, the detection and early warning of driver distraction are essential for improving driving safety and pose an important challenge in intelligent transportation systems. Previous studies used traditional machine learning and deep learning transfer models, which have the disadvantages of complicated and time-consuming manual feature engineering, strong subjectivity, and weak generalization performance. In this paper, we propose a fine-grained detection method for driver distraction based on neural architecture search. First, we design an automatic construction algorithm for deep convolutional neural networks based on neural architecture search, which automatically searches for the optimal deep convolutional neural network architecture without human involvement. In addition, we fuse driver-related multisource perception information, use an automatically constructed deep convolutional neural network to extract high-dimensional mapping features, and implement fine-grained detection of various types of driver distraction states. The results on a large-scale multimodal driver distraction dataset demonstrate that our method can efficiently search an optimal deep convolutional neural network, which can quickly converge, and can accurately detect the considered types of driver distraction states, the average detection accuracy reaches 99.7796%; moreover, it has satisfactory robustness.

Nowadays, interest is growing in automating KDD processes. Thanks to the increasing power and decreasing costs of computation devices, the search for the best features and model parameters can be conducted with different meta-heuristics. Thus, researchers can focus on other important tasks like data wrangling or feature engineering. This article details a comparative study of a GAparsimony R Package with six model complexity metrics. The objective was to identify an adequate model complexity measure for searching for accurate parsimonious solutions by combining feature selection, hyperparameter optimization, and parsimonious evaluation. This study also includes a regression code example to address some recommended precautions and considerations to find robust parsimonious models. This code can be easily adapted to other problems, databases, or algorithms.

Key Message: Convolutional Neural Networks (CNNs) can perform similarly or better than standard genomic prediction methods when sufficient genetic, environmental, and management data are provided. Abstract: Predicting phenotypes from genetic (G), environmental (E), and management (M) conditions is a long-standing challenge with implications to agriculture, medicine, and conservation. Most methods reduce the factors in a dataset (feature engineering) in a subjective and potentially oversimplified manner. Deep neural networks such as Multilayer Perceptrons (MPL) and Convolutional Neural Networks (CNN) can overcome this by allowing the data itself to determine which factors are most important. CNN models were developed for predicting agronomic yield from a combination of replicated trials and historical yield survey data. The results were more accurate than standard methods when tested on held-out G, E, and M data (r = 0.50 vs. r = 0.43), and performed slightly worse than standard methods when only G was held out (r = 0.74 vs. r = 0.80). Pre-training on historical data increased accuracy compared to trial data alone. Saliency map analysis indicated the CNN has “learned” to prioritize many factors of known agricultural importance.

Machine learning models are a promising, yet underutilized tool within the mindfulness field. Accordingly, this work aimed to provide a practical introduction to key machine learning concepts through an illustrative investigation of the association between at-home mindfulness exercise compliance and stress reduction. To further interrogate the currently inconclusive nature of the compliance-outcome association within the mindfulness literature, the illustrative example leveraged a suite of machine learning techniques to highlight the unique affordances and perspectives of the predictive framework. Foundational information regarding facets of the machine learning analytical process, including model types, data preprocessing, feature engineering, validation, performance evaluation, and model introspection, was presented. With emphasis on providing details and justifications regarding modeling decisions along the way, the work systematically applied these introduced concepts to a real-world data example. This permitted an opportunity to build, introspect, and derive insight from a model tasked to explore dynamics underlying patient compliance to mindfulness exercises within a web-based delivery setting. The constructed machine learning models suggested a moderate correlation of compliance with post-intervention reliable change in stress (r = 0.349 ± 0.018). Model introspection tools further revealed that a combination of both high consistency and high overall average compliance predicts a trend toward greater reduction in self-reported stress. Results of the illustrative study suggested that compliance, in pattern and absolute magnitude, is a significant contributor to online mindfulness therapy outcomes. Moreover, modeling efforts implicate machine learning as a uniquely beneficial paradigm with which to explore nuanced questions in the mindfulness research space.

Identification of individuals based on transit modes is of great importance in user tracking systems. However, identifying users in real-life studies is not trivial owing to the following challenges: 1) activity data containing both temporal and spatial context are high-order and sparse; 2) traditional two-step classifiers depend on trajectory patterns as input features, which limits accuracy especially in the case of scattered and diverse data; 3) in some cases, there are few positive instances and they are difficult to detect. Therefore, approaches involving statistics-based or trajectory-based features do not work effectively. Deep learning methods also suffer from the problem of how to represent trajectory vectors for user classification. Here, we propose a novel end-to-end scenario-based deep learning method to address these challenges, based on the observation that individuals may visit the same place for different reasons. We first define a scenario using critical places and related trajectories. Next, we embed scenarios via path-based or graph-based approaches using extended embedding techniques. Finally, a two-level convolution neural network is constructed for the classification. Our model is applied to the problem of detection of addicts using transit records directly without feature engineering, based on real-life data collected from mobile devices. Based on constructed scenario with dense trajectories, our model outperforms classical classification approaches, anomaly detection methods, state-of-the-art sequential deep learning models, and graph neural networks. Moreover, we provide statistical analyses and intuitiveexplanations to interpret the characteristics of resident and addict mobility. Our method could be generalized to other trajectory-related tasks involving scattered and diverse data.

Human activity recognition (HAR) has become a significant area of research in human behavior analysis, human–computer interaction, and pervasive computing. Recently, deep learning (DL)-based methods have been applied successfully to time-series data generated from smartphones and wearable sensors to predict various activities of humans. Even though DL-based approaches performed very well in activity recognition, they are still facing challenges in handling time series data. Several issues persist with time-series data, such as difficulties in feature extraction, heavily biased data, etc. Moreover, most of the HAR approaches rely on manual feature engineering. In this paper, to design a robust classification model for HAR using wearable sensor data, a hybrid of convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) is used. The proposed multibranch CNN-BiLSTM network does automatic feature extraction from the raw sensor data with minimal data pre-processing. The use of CNN and BiLSTM makes the model capable of learning local features as well as long-term dependencies in sequential data. The different filter sizes used in the proposed model can capture various temporal local dependencies and thus helps to improve the feature extraction process. To evaluate the model performance, three benchmark datasets, i.e., WISDM, UCI-HAR, and PAMAP2, are utilized. The proposed model has achieved 96.05%, 96.37%, and 94.29% accuracies on WISDM, UCI-HAR, and PAMAP2 datasets, respectively. The obtained experimental results demonstrate that the proposed model outperforms the other compared approaches.