Lossy text compression reduces data size while preserving core meaning, making it well-suited for tasks like summarization, automated analysis, and digital archives where exact fidelity is less critical. Despite the dominance of transformer-based models in language modeling, the integration of context vectors and lossless entropy coding into Sequence-to-Sequence (Seq2Seq) text generation remains underexplored. A key challenge lies in identifying the most informative context vectors from the encoder output and incorporating entropy coding into the transformer framework to enhance storage efficiency while maintaining high-quality outputs, even in the presence of noisy text. Previous studies have primarily focused on near-lossless token generation, often overlooking space efficiency. In this paper, we introduce TextEconomizer, an encoder-decoder framework paired with a transformer neural network. This framework utilizes its latent representation to reduce variable-sized inputs by 50% to 80%, without prior knowledge of dataset dimensions. Our model achieves competitive compression ratios by incorporating entropy coding, while delivering near-perfect text quality, as assessed by Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), Metric for Evaluation of Translation with Explicit ORdering (METEOR), and semantic similarity scores. Notably, TextEconomizer operates with approximately 153 times fewer parameters than comparable models, achieving a compression ratio of 5.39 ×  without sacrificing semantic quality. Additionally, we evaluate our framework by implementing a Long Short-Term Memory (LSTM)-based autoencoder, commonly used in image compression, and by integrating advanced modules within the transformer architecture as alternatives to conventional techniques. Our autoencoder achieves a state-of-the-art compression ratio of 67 ×  with 196 times fewer parameters, while our modified transformer, LLaMAFormer, achieves a 263-fold reduction in parameters compared to ICAE while maintaining competitive text quality. The TextEconomizer framework significantly surpasses existing transformer-based models in balancing memory efficiency and high-fidelity outputs, marking a breakthrough in lossy compression with optimal space utilization.

The prevalence rate of cyberbullying on social networking sites (SNSs) is a severe issue for online safety. In literature, cyberbullying detection models are mostly modeled on datasets labeled based on standalone tweets. However, bystander roles play an important role in the severity of cyberbullying events. Involving the bystander roles to the detection of the fine-grained cyberbullying could aid in interpreting the intention of short text in standalone tweets. This work extends previous research efforts in the field of cyberbullying detection by considering the whole thread of conversation to capture the bystander roles features. In particular, we introduce the first dataset CYBY24 for thread-based fine-grained cyberbullying detection with bystander roles feature. Furthermore, we tested different detection models with different text representations. Zero-shot classification with LLMs attained less accuracy than the models with access to training data. The proposed step-wise fine-grained cyberbullying detection proved to be the optimal model using BERT and BiLSTM for the first and second classification steps, respectively. We compared the model performance with and without using the proposed bystander roles feature. As a result, the inclusion of bystander roles feature helped the model to improve its predictions.

Network intrusion classification referred to the process of monitoring and analyzing network traffic to identify suspicious activities or attacks. In this work, author proposed a novel approach to classify network intrusion by utilizing deep reinforcement learning (DRL), integrating a reservoir computing approach Echo State Network (ESN). A DRL-based approach improved upon traditional deep learning by adapting dynamically to novel/unknown and evolving attack patterns. Unlike static models, DRL continuously learned optimal strategies through interaction with the environment, allowing for better detection of previously unseen threats in real-time. To address the class imbalance often encountered in network intrusion datasets, we evaluated the performance of several advanced data balancing techniques, including Borderline-SMOTE, SMOTE-ENN, ADYSN, and K-means SMOTE. The findings demonstrated that the K-means-based data balancing method outperformed other techniques, resulting in the most robust performance across various metrics. Author conducted multi-dataset validation on benchmark datasets like NF-BoT-IoT, NF-UNSW-NB15, NF-ToN-IoT, NF-ToN-IoT-v2, NF-CSE-CIC-IDS2018 and NF-UNSW-NB15-v3 to ensure robustness across different network flow data. For adaptive modeling testing, author excluded some attack types from training data and included them in testing data (e.g., DoS, Backdoor attacks were excluded from the training data but included in the testing data (see Table 3)). The proposed approach enhanced the accuracy and reliability of intrusion detection, making it a viable solution for securing modern network infrastructures. The source code of this work is available in this Github repository (https://github.com/codewithkhurshed/DRLZDNIDS).

Ensemble clustering leverages multiple methods to identify diverse patterns and, instead of depending on a singular approach, generates a more dependable and accurate clustering solution. This methodology mitigates bias and noise in intricate, high-dimensional data, allowing the grouping of biological and genomic big data. Component-based ensemble clustering divides data into subsets, applies several algorithms, and then aggregates the outcomes to increase performance. This method analyzes each data subset independently, facilitating the recognition of various patterns while minimizing noise and bias. This paper proposes two novel clustering methods that integrate multiple algorithms, including Agglomerative Hierarchical Clustering (AHC), K-Means Clustering, Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), Ordering Points to Identify the Clustering Structure (OPTICS), Improved Density-Based Spatial Clustering of Applications with Noise (IDBSCAN), and Density-Based Spatial Clustering of Applications with Noise Plus Plus (DBSCAN++). The second method, termed Ensemble Clustering with Each Subset (ECES), employs both ‘with-replacement’ and ‘without-replacement’ techniques to increase variety, minimize redundancy, and improve generalization. The key distinction resides in the ensemble step of the second strategy, which divides datasets into equal subsets to ensure fairness and comparability. This ensures fairness, comparability, and controlled diversity within the ensemble, reducing bias, redundancy, and overlap.

Imbalanced intrusion classification is a complex and challenging task as there are few number of instances/intrusions generally considered as minority instances/intrusions in the imbalanced intrusion datasets. Data sampling methods such as over-sampling and under-sampling methods are commonly applied for dealing with imbalanced intrusion data. In over-sampling, synthetic minority instances are generated e.g. SMOTE (Synthetic Minority Over-sampling Technique) and on the contrary, under-sampling methods remove the majority-class instances to create balanced data e.g. random under-sampling. Both over-sampling and under-sampling methods have the disadvantages as over-sampling technique creates overfitting and under-sampling technique ignores a large portion of the data. Ensemble learning in supervised machine learning is also a common technique for handling imbalanced data. Random Forest and Bagging techniques address the overfitting problem, and Boosting (AdaBoost) gives more attention to the minority-class instances in its iterations. In this paper, we have proposed a method for selecting the most informative instances that represent the overall dataset. We have applied both over-sampling and under-sampling techniques to balance the data by employing the majority and minority informative instances. We have used Random Forest, Bagging, and Boosting (AdaBoost) algorithms and have compared their performances. We have used decision tree (C4.5) as the base classifier of Random Forest and AdaBoost classifiers and naïve Bayes classifier as the base classifier of the Bagging model. The proposed method Adaptive TreeHive addresses both the issues of imbalanced ratio and high dimensionality, resulting in reduced computational power and execution time requirements. We have evaluated the proposed Adaptive TreeHive method using five large-scale public benchmark datasets. The experimental results, compared to data balancing methods such as under-sampling and over-sampling, exhibit superior performance of the Adaptive TreeHive with accuracy rates of 99.96%, 85.65%, 99.83%, 99.77%, and 95.54% on the NSL-KDD, UNSW-NB15, CIC-IDS2017, CSE-CIC-IDS2018, and CICDDoS2019 datasets, respectively, establishing the Adaptive TreeHive as a superior performer compared to the traditional ensemble classifiers.

In this paper, we have studied three well-known classification algorithms: (1) Decision Tree, (2) Naïve Bayesian Classifier, and (3) Naïve Bayesian Tree for supervised learning in machine learning. A decision tree is a top-down recursive divide-and-conquer algorithm, a hierarchical, flowchart-style structure where internal nodes perform attribute-based tests, branches indicate the outcomes of those tests, and leaf nodes represent class labels. Decision trees often face overfitting issues during classification, resulting in overly large trees with irrelevant branches. To address this, pruning techniques eliminate redundant branches, reducing unnecessary tree complexity. The Naïve Bayesian (NB) classifier is a probabilistic model based on Bayes' Theorem that assumes feature independence given the class label. A Naïve Bayesian Tree (NBTree) is a hybrid model that merges decision tree with Naïve Bayes classifier, offering a strong predictive model. It handles attribute interactions through decision splits at internal nodes and uses probabilistic classification at the leaves. This structure enhances accuracy in complex or overlapping data and provides better resistance to overfitting compared to decision trees. This paper analyses the performance of decision tree, naïve Bayesian Classifier, and naïve Bayesian tree on 10 benchmark classification datasets from the UCI Machine Learning Repository using accuracy, precision, recall, and F1-score. We have applied the pruning technique with the decision tree and naïve Bayesian tree to address repetition and replication. The study identifies various patterns and insights from experimental analysis to support the classification task in supervised learning.

The Internet of Things (IoT) has experienced substantial growth in recent years, leading to a significant increase in the number of Internet-connected devices. This rapid expansion has raised concerns regarding the escalating frequency of cyber-attacks. So, it is of utmost importance to have an effective and reliable intrusion detection system (IDS) as part of a comprehensive defense strategy. Recent studies have demonstrated that the performance of IDS can be significantly improved by utilizing machine learning techniques. However, existing centralized techniques involve data sharing, which can increase the computational load and raise privacy concerns. In this paper, we utilize Federated Learning (FL), a distributed machine learning approach that minimizes data sharing and enhances privacy and performance. Additionally, an IDS requires a comprehensive and heterogeneous dataset with sufficient training data to achieve optimal performance. The scarcity of attack data creates an imbalance in the dataset, negatively impacting the model’s effectiveness. To address this problem, We use Generative Adversarial Networks (GANs) to augment the rare class data. Subsequently, We utilize the ANOVA feature selection method to down-sample the training dataset and obtain a rebalanced, low-dimensional dataset. In this article, we propose a novel approach to Intrusion Detection Systems (IDS) by combining Generative Adversarial Networks (GANs) and Federated Learning (FL). The proposed system, FLGAN-IDS, was experimentally evaluated using the NSL-KDD and CIC-IDS2017 datasets for binary and multi-class classification. The findings of this study demonstrate that the FLGAN-IDS model outperformed alternative methods in terms of accuracy, precision, and efficiency, exhibiting remarkable recall and F1-score metrics.

Brain tumors are one of the most common diseases that lead to early death if not diagnosed at an early stage. Traditional diagnostic approaches are extremely time-consuming and prone to errors. In this context, computer vision-based approaches have emerged as an effective tool for accurate brain tumor classification. While some of the existing solutions demonstrate noteworthy accuracy, the models become infeasible to deploy in areas where computational resources are limited. This research addresses the need for accurate and fast classification of brain tumors with a priority of deploying the model in technologically underdeveloped regions. The research presents a novel architecture for precise brain tumor classification fusing pretrained ResNet152V2 and modified VGG16 models. The proposed architecture undergoes a diligent fine-tuning process that ensures fine gradients are preserved in deep neural networks, which are essential for effective brain tumor classification. The proposed solution incorporates various image processing techniques to improve image quality and achieves an astounding accuracy of 98.36 in Figshare and Kaggle datasets respectively. This architecture stands out for having a streamlined profile, with only 2.8 million trainable parameters. We have leveraged 8-bit quantization to produce a model of size 73.881 MB, significantly reducing it from the previous size of 289.45 MB, ensuring smooth deployment in edge devices even in resource-constrained areas. Additionally, the use of Grad-CAM improves the interpretability of the model, offering insightful information regarding its decision-making process. Owing to its high discriminative ability, this model can be a reliable option for accurate brain tumor classification.

Emotion recognition is becoming increasingly relevant in many domains, including healthcare, human-computer interaction, and affective computing. Accurate emotion recognition is essential for improving user experience, conducting mental health assessments, and developing adaptive systems. In our study, we provide a novel multimodal approach that combines visual and thermal data to improve emotion identification. Our system effectively extracts and merges features from both modalities by employing CNN architectures (VGG16, MobileNet, and ResNet50 for visual data and VGG16, MobileNet for thermal data). The fused feature set is classified using support vector machines (SVMs) and decision trees, with SVM accuracy of 97.70% and Decision Tree accuracy of 93.72%. Furthermore, SVM produced F1-scores of 97.31%, greatly exceeding the Decision Tree model with 88.40%. Class imbalance was reduced using strategic data augmentation, which improved the model’s performance even further. This study emphasizes the efficiency of multimodal systems by proving that mixing visual and thermal data improves the accuracy and resilience of emotion recognition. The suggested system’s versatility and efficiency make it an important tool for real-world applications including healthcare and human-computer interaction.

Dynamic feature selection is critical for improving the flexibility and efficiency of predictive models in machine learning, particularly when dealing with sequential data streams. In this study, we have introduced a novel framework, the Recurrent Feature Importance Network (ReFIN), which integrates Recurrent Neural Networks (RNNs) with decision tree principles to dynamically select and adjust feature importance over time. This approach begins by randomly selecting subsets of features and evaluating their performance using a Random Forest Classifier with attribute bagging. The performance of each subset is assessed, and the best-performing feature sets are clustered into categories of Good, Average, and Poor importance. These selected features are then utilized to train the ReFIN model, which leverages an LSTM-based RNN architecture to continuously update feature importance based on evolving data contexts. Through systematic iteration, the proposed method adjusts feature importance dynamically, ensuring that the most relevant features are prioritized while less important ones are down-weighted or replaced as data patterns shift. This adaptability is particularly effective in handling concept drift, a common challenge in data stream classification, where data distributions change over time. Experimental results demonstrate that ReFIN significantly enhances model accuracy and efficiency by optimizing feature selection and improving the interpretability of the classification process in dynamic environments.

Individuals with visual impairments face challenges accessing written information. Braille is one of many solutions that only requires the reader to sense the depth of the paper with their hand to comprehend written information without the need to look at the text. However, learning Braille, especially for those losing sight later in life, presents difficulties. This research introduces the BrailleSense system, a technological solution designed to assist visually impaired individuals in learning and utilizing the Braille system effectively. The system features a virtual prototype of hand gloves equipped with a camera, aiming to alleviate challenges associated with Braille pattern memorization. Key contributions include the development of a custom lightweight Convolutional Neural Network (CNN) model for Braille pattern classification coined as the BrailleNet. This model is then deployed on a Raspberry Pi to investigate the feasibility of working with resource-limited portable devices, BrailleNet achieves an impressive accuracy of 97.44% under real-world constraints. The research outlines the conceptual design through a 3D model of the gloves, addressing spatial allocation. Acknowledging challenges in user comfort and alignment, BrailleSense presents a pioneering step towards empowering visually impaired individuals, enhancing literacy, and fostering independence. The dataset and code for the BrailleSense system are available on GitHub - https://github.com/faiyazabdullah/BrailleNet

An optimized educational community is a must in this modern era. The intersection of educational activities and the transformative potentials of Educational Data Mining (EDM) should be traversed, highlighting the reasoning behind the importance of EDM. Prior prediction of how a student stands academically, can facilitate them towards a much safer approach with their life decisions. This study uses the vast power and analytical domain of EDM, combining it with machine learning models, upholding an accurate prediction of students' academic performance. The study consists of a dataset containing academic, demographic and social data of undergraduate students. The paper aims to analyze comprehensively the features that act behind academic performance. Lastly, it compares the impact of non-academic data separately on a student's performance and with academic data as well. Traditional machine learning algorithms perform quite well in general, with SVM giving a best accuracy of around 95% with academic data, while training and testing the model without academic data still gives a good performance of 93%. The hierarchical tree from Decision Tree visualizes the key features, which include past results, family members' qualification levels and their jobs, hobbies of the student, commute time, and more.

This research was conducted with a view to develop personalized awareness tools which can play an important role against drug abuse by visualizing the potential impacts of addiction. A novel dataset was prepared for this purpose consisting of before-and-after images of individuals affected by drug addiction. Two popular GAN architectures - Pix2Pix and CycleGAN were developed in this research, utilizing transfer learning to enhance training efficiency. A comparative analysis revealed that while Pix2Pix is adept at incorporating distinct addiction-related features into the generated images, it often fails to maintain the facial structure of the individuals. On the contrary, CycleGAN maintains structural integrity but it is less precise in transferring addiction-specific characteristics. These insights reveal both the strengths and weaknesses of using these technologies for personalized medical and awareness applications.

Educational Data Mining (EDM) is used to ameliorate the teaching and learning process by analyzing and classifying data that can be applied to predict the students’ academic performance, and students’ dropout rate, as well as instructors’ performance. The prediction of student performance is complicated by the vast and diverse range of variables from academic records to behavioral and health metrics. In this paper, we have introduced a new Adaptive Feature Selection Algorithm (AFSA) by amalgamating an ensemble approach for initial feature ranking with normalized mean ranking from five distinct methods to enhance robustness. The proposed method iteratively selects the best features by adjusting its threshold based on each feature’s rank to ensure significant contributions to model accuracy and also effectively reduces dataset complexity. We have tested the performance of the proposed feature selection algorithm using five machine learning classifiers: Logistic Regression (LR), K-Nearest Neighbour (KNN), Support Vector Machine (SVM), Naïve Bayes (NB) classifier, and Decision Tree (DT) classifier on four student performance datasets. The experimental results highlight the proposed method significantly decreases feature count by an average feature reduction factor of 5.7, significantly streamlining datasets while maintaining competitive cross-validation accuracy, marking it as a valuable tool in the field of educational data analytics.

Decision tree is one of the most popular classifiers in supervised learning to address classification tasks. It is easy to build and easy to understand. It is a divide-and-conquer algorithm and provides a clear path for making decisions. Decision tree can be used for both regression and classification problems. It finds the best splitting features as a root node and create sub-datasets. A leaf node of decision tree represents a decision where all and majority of the instances of a sub-dataset belongs to a particular decision or class-value. The commonly used algorithms are ID3 (Iterative Dichotomiser 3), C4.5 (improvement over ID3 algorithm), and CART (Classification And Regression Tree). Decision tree can be easily scalable and apply for Big Data mining. In this paper, we have proposed a new pruning-based tree ensemble method where each tree is build from the each input feature of the dataset. Therefore, $N$ number of trees will be created for $N$ number of features. Then, we apply post pruning technique on each tree and select the best trees according to their cost-complexity. We have considered the size and accuracy of the tree to select in the ensemble process. The proposed approach surpasses the performance of the C4.5 classifier by almost 11 % for the selected benchmark datasets taken from the UCI machine learning repository.

Medical image classification is critical in clinical decision-making, requiring efficient and precise evaluation of images. Convolutional Neural Networks (CNNs) have significantly contributed to this field, yet their resource-intensive nature poses challenges. This study aims to analyse the effectiveness of vision transformers with registers for medical image classification on the MedMNIST dataset, comparing them with other state-of-the-art classifiers. We propose an E-MedViTR model that boosts the performance of the vision transformer with registers by adding an enhanced classifier head. Our suggested model surpasses existing counterparts, achieving the highest F-1 scores in multi-class classification, demonstrating its effectiveness in colon pathology image classification. The model scores an accuracy of 85.80% on the PathMNIST, one of the sub-datasets of the MedMNIST dataset. Research in this field could pave the way for more personalized, accurate, and efficient diagnostic procedures in medical imaging, benefiting both clinicians and patients.

In the contemporary landscape of business intelligence and market analysis, customer segmentation serves as a pivotal tool for understanding consumer behavior and preferences. This paper delves into the application of advanced machine learning techniques, specifically K-Modes clustering and ensemble learning with AdaBoost, for the purpose of customer segmentation and classification. The utilization of K-Modes clustering, an extension of the K-Means algorithm tailored for categorical data, facilitates the identification of distinct groups within a heterogeneous customer base. By incorporating categorical variables, K-Modes accommodates the inherent diversity in customer attributes such as demographic information, purchase history, and product preferences. Furthermore, this research integrates ensemble learning techniques, particularly AdaBoost, to enhance the accuracy and robustness of the segmentation process. Through a comprehensive empirical analysis, conducted on a real-world dataset sourced from Kaggle, the proposed methodology demonstrates superior performance compared to traditional clustering approaches. The experimental results showcase the effectiveness of K-Modes clustering combined with AdaBoost ensemble learning in accurately segmenting customers into meaningful groups, thereby enabling businesses to gain deeper insights into consumer behavior and preferences.

Multi-class classification presents a significant challenge in supervised machine learning, and it is frequently applied across various real-world domains. Random Forest (RF) stands out as a popular ensemble technique for multi-class classification tasks that utilizes attribute bagging to construct random trees and employs a majority-voting scheme for classification. However, the existing Random Forest approach often faces limitations in accuracy and efficiency, particularly for more critical tasks. In this study, we have addressed this gap by introducing a novel methodology that integrates the K-Nearest Neighbor (KNN) classifier and Decision Trees into the Random Forest framework to enhance performance. The proposed method has involved training a K-Nearest Neighbor and a Decision Tree (DT) classifier on the extracted features. The output probabilities from these classifiers have been used as meta-features, forming an enriched feature vector fed into a Random Forest as the meta-learner for final classification. Through extensive experimentation on 5 datasets, the proposed approach has demonstrated superior performance in accuracy and efficiency compared to the traditional method. By integrating Random Forest and K-Nearest Neighbor techniques, this research has offered a versatile solution applicable across diverse domains where precise multi-class classification is essential.

Alzheimer’s disease (AD) is a progressive neurological disorder that significantly impairs cognitive functions, particularly memory and thinking skills. The presence of AD in millions of individuals worldwide constitutes a substantial global health challenge. Timely and accurate diagnosis of AD is critical for effective management and improved patient outcomes. This study introduces ViTAD, an innovative method for classifying five stages of AD from brain MRI images, leveraging a Vision Transformer (ViT) model. The proposed model modifies Google’s ViT architecture, incorporating fine-tuned hyperparameters and additional layers to enhance its performance for AD stage detection. The dataset comprises 1,296 brain MRI images from the ADNI dataset, covering five stages of AD: Cognitively Normal (CN), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), Mild Cognitive Impairment (MCI), and Alzheimer’s Disease (AD). Our preprocessing pipeline includes grayscale to RGB conversion, image cropping, and the application of a Laplacian sharpening filter to enhance image clarity. Data augmentation was performed using horizontal/vertical flips, zoom, and rotation to ensure model robustness. We allocated 85% of the dataset for training and 15% for testing. Upon training the model for 20 epochs with a learning rate of 0.0001, ViTAD achieved a remarkable 99.98% accuracy, with 100% precision and an F1-score of 1.00. ViTAD’s superior performance in the multi-class classification task outperforms several conventional CNN-based models such as DenseNet and EfficientNet, which struggled with the 5-class AD detection task. Additionally, ViTAD demonstrated high efficiency, achieving optimal accuracy within only 8 epochs, far surpassing traditional CNN models in speed and accuracy. These findings highlight the significant potential of ViTAD as an automated, accurate, and efficient tool for early diagnosis of AD, offering valuable support in clinical settings.

A Network Intrusion Detection System (NIDS) serves as a sentinel for safeguarding data integrity. It watches over computer networks, looking out for and stopping threats that can sneak past normal defenses like malware and hackers. Deep Learning (DL) techniques offer a promising avenue for analyzing raw network data to uncover subtle patterns indicative of intrusion attempts. In this study, we address a critical research gap by developing a Deep Neural Network (DNN) model tailored for efficient detection of stealthy and polymorphic variants while mitigating false positives. Leveraging the NF-ToN-loT dataset, the proposed model achieves impressive performance metrics on test data, with an accuracy of 0.99, precision of 0.98, recall of 0.99, and F1-score of 0.99. To comprehensively assess the robustness of the proposed model, we use a multi-dataset validation strategy. The model is retrained and evaluated on established benchmark datasets, including NF-BoT-loT, NF-UNSW-NB15, and NF-UNSW-NB15-v2, demonstrating exceptional performance. Furthermore, to ensure the significance of our contribution, we compare our model against previously well-established architectures such as CNN+BiLSTM, DNN, GRU+RNN, and CNN+LSTM. Utilizing the NF-ToN-loT dataset as a common ground, the proposed model demonstrably outperforms these prior models, highlighting its efficacy and advancement in the field. Additionally, we conduct an ablation study to dissect the components of the DNN model, shedding light on their individual contributions towards detecting malware traffic and offering insights for optimizing future NIDS models in the cybersecurity domain.