Convolutional neural networks often contain redundant feature maps that increase storage and inference cost. This paper presents a loss-aware feature-map pruning framework using multi-armed bandits. Feature-map pruning is structured because it removes complete convolutional output channels and their producing filters rather than isolated scalar weights. Each candidate feature map is treated as an arm. At each play time, one map is temporarily masked and evaluated on a sampled mini-batch; the map is then restored and the observed loss change is converted into a safe-removal reward. After a fixed play budget, candidate maps are ranked by learned scores and the top-k maps are permanently removed with their filters, biases and corresponding next-layer input-channel kernels. The study evaluates UCB1 and Thompson Sampling, compares them with direct/oracle-style evaluation on LeNet/MNIST, and extends the evaluation to MNIST, CIFAR-10, CIFAR-100, SVHN, CUB-200-2011 and Oxford Flowers 102. Results show that UCB1 and Thompson Sampling preserve accuracy close to unpruned models while removing feature maps and reducing convolutional computation. Friedman and Nemenyi tests show that UCB1 obtains the highest mean rank, followed by Thompson Sampling; both significantly outperform greedy and magnitude-based pruning while remaining statistically comparable to the original unpruned model.

Deep neural networks often contain redundant hidden units. Removing individual weights can reduce parameter count, but unstructured sparsity is not always easy to exploit in standard dense implementations. This paper develops a structured pruning framework in which complete neurons are removed using multi-armed bandit (MAB) algorithms. Each candidate neuron is treated as an arm; pulling an arm temporarily masks that neuron, measures the change in loss on a sampled mini-batch, restores the neuron, and updates an estimate of its safe-removal reward. The framework supports stochastic policies, including Epsilon-Greedy, Softmax, UCB1 and Thompson Sampling, and multiplicative-weight policies, including Hedge-style multiplicative weights and EXP3. We evaluate the method on tabular classification, tabular regression and deep neural-network benchmarks covering image, text and reasoning tasks. Statistical comparisons using the Friedman test followed by the Nemenyi post-hoc test show significant differences between methods. On tabular classification tasks, UCB1 obtains the highest mean rank among pruning policies and improves on the unpruned neural network. On regression tasks, UCB1 obtains the highest mean rank and is statistically competitive with, or superior to, several standard regression models according to R^2. On deep-learning tasks, UCB1 and Thompson Sampling obtain the strongest ranks, and several MAB policies significantly outperform the unpruned model, magnitude-based neuron pruning and greedy activation-variation pruning. The results show that MAB-based neuron pruning is an effective and computationally practical approach for structured model reduction.

The increasing integration of cyber-physical systems in critical infrastructure like water treatment plants has heightened their vulnerability to sophisticated attacks, necessitating robust anomaly detection mechanisms. This paper presents a Temporal-Enhanced LSTM Autoencoder (TE-LSTM-AE), a novel two-stage deep learning model designed to detect cyber-physical attacks by learning complex temporal patterns and cross-sensor correlations. Our group-aware architecture processes sensor data based on functional characteristics (e.g., hydraulic performance, water quality) in dedicated autoencoders before a fusion stage learns critical inter-group relationships. Evaluated on the Secure Water Treatment (SWaT) testbed, the model achieved state-of-the-art performance with a precision of 0.9910, recall of 0.9244, F1-score of 0.9565, and a near-perfect ROC-AUC of 0.9971. Detailed analysis demonstrates the model's high sensitivity to disruptions in hydraulic correlations and its unique capability to identify a concentrated multi-stage attack campaign. The results establish that a physics-informed, group-based approach is superior for achieving high-fidelity, interpretable, and robust anomaly detection in complex industrial environments.

Anomaly detection in Industrial Control Systems (ICS) is critical for ensuring the stability and security of critical infrastructure, such as water treatment plants, where undetected faults can lead to catastrophic failures. Traditional methods, including standalone autoencoders and Long Short-Term Memory (LSTM) networks, often struggle to capture both temporal dependencies and long-term behavioural trends in complex, high-dimensional sensor data. While hybrid models like LSTM-autoencoders (LSTM-AEs) have improved anomaly detection by integrating feature extraction and sequential learning, they remain limited in identifying slow-developing faults and subtle operational shifts. To address these challenges, this paper introduces the Trend-Enhanced LSTM Autoencoder (TE-LSTM AE), a novel dual-stage framework that cascades two LSTM-AE models. The first stage extracts latent representations of input data, while the second uses these extracted trends as its input features to detect anomalies by leveraging both fine-grained temporal patterns and broader system trends. This architecture enhances sensitivity to gradual anomalies, such as sensor degradation, while maintaining robustness against abrupt deviations. Evaluated on Secure Water Treatment (SWaT) dataset, TE-LSTM AE achieved state-of-the-art results, with an F1 Score of 0.9732, an ROC AUC of 0.98, and a Precision-Recall AUC of 0.90, outperforming baseline models like LSTM, GRU, LSTM-AE and GRU-AE, and demonstrating an ability to detect subtle deviations while maintaining high precision.

Human emotion classification without bias and unfairness is challenging because most existing image-based methods are directly or indirectly affected by subjectivity. Therefore, we propose an EEG (Electroencephalogram) based model for an accurate emotion classification without the effect of subjectivity. The captured EEG signals are converted into Delta, Theta, Alpha, Beta, and Gama frequency bands. As emotions change, the frequency bands change and provide unique patterns for each emotion irrespective of different persons. With this observation, the statical features, namely, mean, standard deviation, variance, and kurtosis, and frequency-based features, namely, Power Spectral Density (PSD) and Petrosian Fractal Dimension (PFD) are extracted. To integrate the strength of spatial and frequency-based features, the features are supplied to quadratic discriminative analysis for the final classification. The experiments on the benchmark datasets, DEAP and SEED-IV, achieve 99.40% and 91.97% accuracy, respectively. A comparison with state-of-the-art methods shows that the method performs very well on some datasets.

ObjectiveCardiovascular disease (CVD) is the leading cause of premature death in the United Kingdom with one type, coronary artery disease, killing more than two times as many women as breast cancer. Recently, researchers have noted that breast arterial calcification (BAC), which is regularly observed as an incidental finding on mammograms, could be used to risk-stratify women for CVD. However, identifying BAC is known to be a tedious, expensive and time-consuming process. Thus, this paper investigates deep learning models for BAC classification, object detection and segmentation.MethodologyA data set, annotated under the guidance of two consultant radiologists, was created using data augmentation. This was used to evaluate several alternative deep learning models.ResultsA modified ResNet22 classification network achieved a test accuracy of 80%, indicating that this method could be used as a flag for the presence or absence of BAC. We also used this network for feature extraction in a YOLOv4 BAC object detection network. Despite improving on a recent similar study, this latter network performed poorly with very low average precision scores at several thresholds. More promising was our DeepLabv3+-based BAC segmentation network, which reached similar high global accuracy scores to three recent studies and a BFScore of over 70% specifically for BAC. It also performed satisfactorily on an unseen data set.ConclusionsThese results show the potential for using classification and segmentation models as part of a pipeline for detecting BAC.

The interactive behaviour of pigs is an important determinant of their social development and overall well-being. Manual observation and identification of contact behaviour can be time-consuming and potentially subjective. This study presents a new method for the dynamic detection of pig head to rear interaction using the Vision Transformer (ViT). The ViT model achieved a high accuracy in detecting and classifying specific interaction behaviour as trained on the pig contact datasets, capturing interaction behaviour. The model's ability to recognize contextual spatial data enables strong detection even in complex contexts, due to the use of Gaussian Error Linear Unit (GELU) an activation function responsible for introduction of non-linear data to the model and Multi Head Attention feature that ensures all relevant details contained in a data are captured in Vision Transformer. The method provides an efficient method for monitoring swine behaviour for instance, contact between pigs, facilitating better livestock management and livestock welfare. The ViT can represent a significant improvement on current automated behaviour detection, opening new possibilities for accurate animal design and animal behaviour assessment with an accuracy and F1 score of 82.8% and 82.7%, respectively, while we have an AUC of 85%.

Purpose This paper aims to introduce a new method for pre-processing data to reduce bias in pricing models without compromising profit. The authors develop a straightforward yet impactful technique, called Mean Penalty Bias Reduction, which effectively eliminates unfair bias while preserving overall revenue. This method applies a transformation that ensures the revised data meets a desired disparity threshold. This study experiments with this approach to verify that the revised data satisfies the criteria for a widely accepted fairness measure. Design/methodology/approach The paper reviews existing work and identifies an appropriate measure of fairness, known as disparate impact, which is adapted for pricing. A pre processing transformation that reduces disparate impact is formally derived and proved to maintain profitability. The new transformation is then used to develop an algorithm which is evaluated on two data sets: a publicly available insurance data set and a synthetically created airlines data set with introduced bias. The insurance data set offers real-world applicability, and the synthetic airlines data set allows controlled experimentation with known biases, ensuring the robustness of the findings in different scenarios. Findings This paper offers empirical insights into the critical importance of fairness in pricing and the need to ensure that machine learning models are not biased against protected characteristics. It introduces a new methodology that businesses can use to reduce bias in pricing models without compromising profitability. This approach highlights the growing relevance of ethical considerations in today’s world, demonstrating how companies can achieve equitable outcomes while maintaining their financial objectives. Research limitations/implications Because the suggested methodology aims to maintain revenue, it reduces, rather than completely eliminates, bias while removing disparate impact. Naturally, pre-processing methods alone are insufficient; continuous monitoring and assessment are essential to ensuring a model’s fairness over time. Practical implications The practical implication of this methodology is that businesses can achieve a balance between fairness and profitability in their pricing models. By reducing bias and removing disparate impact, companies can foster more equitable customer relationships and comply with regulatory standards. However, because pre-processing methods alone are not enough, businesses must implement ongoing monitoring and assessment to ensure the continued fairness of their models. This approach not only mitigates ethical and legal risks but also enhances the company’s reputation and trustworthiness in the eyes of consumers and stakeholders. Social implications Implementing this methodology has significant social implications by promoting fairness and reducing discrimination in pricing models. By addressing bias against protected characteristics, companies can contribute to greater social equity and inclusion. This approach helps to ensure that all customers are treated fairly, regardless of their background, fostering trust and loyalty. In addition, as businesses adopt fairer practices, societal awareness and demand for ethical standards in machine learning and data usage will likely increase, encouraging a broader cultural shift towards equity and justice in various industries. Ultimately, this contributes to a more just and equitable society. Originality/value This paper addresses a recognized need to explore how fairness in pricing can be identified and reduced without affecting the revenue. By investigating methods to mitigate bias in pricing models, this study contributes valuable insights into achieving equitable outcomes. It provides a framework for understanding and addressing disparities, ensuring that pricing strategies do not unfairly disadvantage any group. This research fills an important gap by offering practical solutions and promoting a fairer approach to pricing, ultimately helping businesses and policymakers create more equitable systems that benefit all stakeholders.

Forecasting the academic achievement of students is a critical area of research in educational contexts. This domain's significance stems from its ability to develop efficient mechanisms that enhance academic outcomes and minimize student attrition. In this context, rubric-based progressive learning meticulously provides valuable insights into students' preferences, knowledge, and competencies. This study proposes a recommender model for detecting the Computational Thinking (CT) competencies of programming learners using a rubric and machine learning. A programming rubric was prepared to cover key programming concepts. A quiz conducted afterward was scored as per the rubric designed. Hierarchical clustering was applied to the rubric scores of learners to segment them into four categories according to their learning parameters. The rules were generated as CT competencies using a rule-based classifier—a multiple-layer perceptron neural network, considering cluster categories as labels. The proposed model assists learners and instructors in identifying the learners' learning capabilities and priorities, resulting in improved learner performances.

Phishing involves malicious activity whereby phishers, in the disguise of legitimate entities, obtain illegitimate access to the victims’ personal and private information, usually through emails. Currently, phishing attacks and threats are being handled effectively through the use of the latest phishing email detection solutions. Most current phishing detection systems assume phishing attacks to be in English, though attacks in other languages are growing. In particular, Arabic is a widely used language and therefore represents a vulnerable target. However, there is a significant shortage of corpora that can be used to develop Arabic phishing detection systems. This article presents the development of a new English-Arabic parallel phishing email corpus that has been developed from the anti-phishing share task text (IWSPA-AP 2018). The email content was to be translated, and the task had been allotted to 10 volunteers who had a university background and were English and Arabic language experts. To evaluate the effectiveness of the new corpus, we develop phishing email detection models using Term Frequency–Inverse Document Frequency and Multilayer Perceptron using 1,258 emails in Arabic and English that have equal ratios of legitimate and phishing emails. The experimental findings show that the accuracy reaches 96.82% for the Arabic dataset and 94.63% for the emails in English, providing some assurance of the potential value of the parallel corpus developed.

IoT sensors capture different aspects of the environment and generate high throughput data streams. Besides capturing these data streams and reporting the monitoring information, there is significant potential for adopting deep learning to identify valuable insights for predictive preventive maintenance. One specific class of applications involves using Long Short‐Term Memory Networks (LSTMs) to predict faults happening in the near future. However, despite their remarkable performance, LSTMs can be very opaque. This paper deals with this issue by applying Learning Fuzzy Cognitive Maps (LFCMs) for developing simplified auxiliary models that can provide greater transparency. An LSTM model for predicting faults of industrial bearings based on readings from vibration sensors is developed to evaluate the idea. An LFCM is then used to imitate the performance of the baseline LSTM model. Through static and dynamic analyses, we demonstrate that LFCM can highlight (i) which members in a sequence of readings contribute to the prediction result and (ii) which values could be controlled to prevent possible faults. Moreover, we compare LFCM with state‐of‐the‐art methods reported in the literature, including decision trees and SHAP values. The experiments show that LFCM offers some advantages over these methods. Moreover, LFCM, by conducting a what‐if analysis, could provide more information about the black‐box model. To the best of our knowledge, this is the first time LFCMs have been used to simplify a deep learning model to offer greater explainability.

This paper presents a survey of methods for pruning deep neural networks. It begins by categorising over 150 studies based on the underlying approach used and then focuses on three categories: methods that use magnitude based pruning, methods that utilise clustering to identify redundancy, and methods that use sensitivity analysis to assess the effect of pruning. Some of the key influencing studies within these categories are presented to highlight the underlying approaches and results achieved. Most studies present results which are distributed in the literature as new architectures, algorithms and data sets have developed with time, making comparison across different studied difficult. The paper therefore provides a resource for the community that can be used to quickly compare the results from many different methods on a variety of data sets, and a range of architectures, including AlexNet, ResNet, DenseNet and VGG. The resource is illustrated by comparing the results published for pruning AlexNet and ResNet50 on ImageNet and ResNet56 and VGG16 on the CIFAR10 data to reveal which pruning methods work well in terms of retaining accuracy whilst achieving good compression rates. The paper concludes by identifying some research gaps and promising directions for future research.

IoT sensors and deep learning models can widely be applied for fault prediction. Although deep learning models are considerably more potent than many conventional machine learning models, they are not transparent. This paper first examines different deep learning techniques to carry out univariate time series analysis based on vibration sensors installed on four industrial bearings to predict a fault occurring in a predefined time window. Several recurrent neural networks are used to develop fault prediction models. An empirical evaluation of these models shows that all models perform well; however, hybrid models outperform other models when the time window increases. Then, instance-wise feature selection has been considered to highlight the most contributing features for its outputs regarding any input. In this problem, the main challenge is to propose a trainable feature selection model with the minimum number of selected features whilst its performance is close to the baseline model. This paper develops a novel explainable method called the Gumbel-Sigmoid eXplanator (GSX) to tackle these problems. In a nutshell: (i) we have developed a differentiable and trainable selector, and (ii) we utilize regularization to control the number of features for each instance flexibly. The proposed method is model agnostic, and empirical evaluations on two datasets show that GSX can not only solve the problems identified with two other state-of-the-art methods but also outperform them in terms of accuracy and run-time.

This paper develops an Evolutionary Elliptical Cost-Sensitive Decision Tree Algorithm (EECSDT) which learns cost-sensitive non-linear decision trees for multiclass problems. EECSDT is developed by formulating the problem as an optimization task in which the objective is to minimize classification cost and where elliptical decision boundaries are adopted instead of axis parallel boundaries. EECSDT is implemented using MOEA, a framework for multi-objective evolutionary algorithms, and evaluated on fourteen data sets. An empirical evaluation with J48, NBTree, MetaCost, and the CostSensitiveClassifier in Weka shows that EECSDT performs better on 11 out of the 14 data sets in terms of accuracy, and 10 out of the 14 data sets in terms of minimizing cost. It also produces smaller trees on 8 out of the 11 datasets for which it achieves higher accuracy than use of axis parallel boundaries.

Purpose This paper aims to describe the use of a meta-learning framework for recommending cost-sensitive classification methods with the aim of answering an important question that arises in machine learning, namely, “Among all the available classification algorithms, and in considering a specific type of data and cost, which is the best algorithm for my problem?” Design/methodology/approach This paper describes the use of a meta-learning framework for recommending cost-sensitive classification methods for the aim of answering an important question that arises in machine learning, namely, “Among all the available classification algorithms, and in considering a specific type of data and cost, which is the best algorithm for my problem?” The framework is based on the idea of applying machine learning techniques to discover knowledge about the performance of different machine learning algorithms. It includes components that repeatedly apply different classification methods on data sets and measures their performance. The characteristics of the data sets, combined with the algorithms and the performance provide the training examples. A decision tree algorithm is applied to the training examples to induce the knowledge, which can then be used to recommend algorithms for new data sets. The paper makes a contribution to both meta-learning and cost-sensitive machine learning approaches. Those both fields are not new, however, building a recommender that recommends the optimal case-sensitive approach for a given data problem is the contribution. The proposed solution is implemented in WEKA and evaluated by applying it on different data sets and comparing the results with existing studies available in the literature. The results show that a developed meta-learning solution produces better results than METAL, a well-known meta-learning system. The developed solution takes the misclassification cost into consideration during the learning process, which is not available in the compared project. Findings The proposed solution is implemented in WEKA and evaluated by applying it to different data sets and comparing the results with existing studies available in the literature. The results show that a developed meta-learning solution produces better results than METAL, a well-known meta-learning system. Originality/value The paper presents a major piece of new information in writing for the first time. Meta-learning work has been done before but this paper presents a new meta-learning framework that is costs sensitive.

Every year, phishing results in losses of billions of dollars and is a major threat to the Internet economy. Phishing attacks are now most often carried out by email. To better comprehend the existing research trend of phishing email detection, several review studies have been performed. However, it is important to assess this issue from different perspectives. None of the surveys have ever comprehensively studied the use of Natural Language Processing (NLP) techniques for detection of phishing except one that shed light on the use of NLP techniques for classification and training purposes, while exploring a few alternatives. To bridge the gap, this study aims to systematically review and synthesise research on the use of NLP for detecting phishing emails. Based on specific predefined criteria, a total of 100 research articles published between 2006 and 2022 were identified and analysed. We study the key research areas in phishing email detection using NLP, machine learning algorithms used in phishing detection email, text features in phishing emails, datasets and resources that have been used in phishing emails, and the evaluation criteria. The findings include that the main research area in phishing detection studies is feature extraction and selection, followed by methods for classifying and optimizing the detection of phishing emails. Amongst the range of classification algorithms, support vector machines (SVMs) are heavily utilised for detecting phishing emails. The most frequently used NLP techniques are found to be TF-IDF and word embeddings. Furthermore, the most commonly used datasets for benchmarking phishing email detection methods is the Nazario phishing corpus. Also, Python is the most commonly used one for phishing email detection. It is expected that the findings of this paper can be helpful for the scientific community, especially in the field of NLP application in cybersecurity problems. This survey also is unique in the sense that it relates works to their openly available tools and resources. The analysis of the presented works revealed that not much work had been performed on Arabic language phishing emails using NLP techniques. Therefore, many open issues are associated with Arabic phishing email detection.

Phishing is a serious form of online fraud made up of spoofed websites that attempt to gain users’ sensitive information by tricking them into believing that they are visiting a legitimate site. Phishing attacks can be detected many ways, including a user's awareness of fraud protection, blacklisting websites, analyzing the suspected characteristics, or comparing them to recent attempts that followed similar patterns. The purpose of this paper is to create classification models using features extracted from websites to study and classify phishing websites. In order to train the system, we use two datasets consisting of 58,645 and 88,647 URLs labeled as “Phishing” or “Legitimate”. A diverse range of machine learning models such as “XGBOOST, Support Vector Machine (SVM), Random Forest (RF), k-nearest neighbor (KNN), Artificial neural network (ANN), Logistic Regression (LR), Decision tree (DT), and Gaussian naïve Bayes (NB)” classifiers are evaluated. ANN provided the best performance with 97.63

s of Keynotes and Invited Presentations Tensor Networks and Their Applications in Image and Time Series Reconstruction and Prediction

The IIP 2020 proceedings present the latest research on intelligent information processing. The papers deal with topics such as machine learning; multi-agent system; recommendation system; social computing; brain computer integration; pattern recognition; and computer vision and image understanding.

The last decade has seen an increase in the attention paid to the development of cost-sensitive learning algorithms that aim to minimize misclassification costs while still maintaining accuracy. Most of this attention has been on cost-sensitive decision tree learning, whereas relatively little attention has been paid to assess if it is possible to develop better cost-sensitive classifiers based on Bayesian networks. Hence, this paper presents EBNO, an algorithm that utilizes Genetic algorithms to learn cost-sensitive Bayesian networks, where genes are utilized to represent the links between the nodes in Bayesian networks and the expected cost is used as a fitness function. An empirical comparison of the new algorithm has been carried out with respect to (a) an algorithm that induces cost-insensitive Bayesian networks to provide a base line, (b) ICET, a well-known algorithm that uses Genetic algorithms to induce cost-sensitive decision trees, (c) use of MetaCost to induce cost-sensitive Bayesian networks via bagging (d) use of AdaBoost to induce cost-sensitive Bayesian networks, and (e) use of XGBoost, a gradient boosting algorithm, to induce cost-sensitive decision trees. An empirical evaluation on 28 data sets reveals that EBNO performs well in comparison with the algorithms that produce single interpretable models and performs just as well as algorithms that use bagging and boosting methods.