Entropy-driven topology mapping framework for robust Bayesian classification

“The Social Dilemma” has triggered significant online discourse, with more than 1.5 million tweets generated globally in the first month of its release and nearly 62% of them reflecting strong opinions on social media regulation. Traditional manual classification of such data is labour-intensive, errorprone, and lacks scalability, limiting the extraction of meaningful insights. To address these issues, a Natural Language Processing (NLP) pipeline is designed that preprocesses the “The Social Dilemma” Tweets dataset, followed by Exploratory Data Analysis (EDA) to identify trends, noise, and dominant language patterns. Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa) with word embeddings is employed for deep contextual feature extraction with SMOTEbalanced features, enhancing semantic understanding beyond surface-level text. The extracted features are used to train and evaluate multiple Machine Learning (ML) classifiers for two parallel targets: sentiment classification and topic classification. Existing approaches such as Decision Tree Classifier (DTC), K-Nearest Neighbor (KNN), and Naïve Bayes Classifier (NBC) are considered for baseline performance comparison. In the proposed methodology, a Deep Neural Network (DNN) feature extractor combined with a Tao Tree Classifier (TTC) ensures higher generalization and robust classification. Hereafter, the proposed system is named “SocialTransformDeepTree.” Sentiments are categorized into Negative, Neutral, and Positive, while topics are classified into Calls for Action, Documentary Recommendation, Emotional Reactions, Irony & Self-Reflection, and Key Quotes & Insights. This hybrid framework not only improves accuracy and computational efficiency but also establishes a scalable approach for analyzing complex social media discussions surrounding documentaries with global societal impact.

Deep neural networks for 3D point cloud classification, such as PointNet, have been demonstrated to be vulnerable to adversarial attacks. Current adversarial defenders often learn to denoise the (attacked) point clouds by reconstruction, and then feed them to the classifiers as input. In contrast to the literature, we propose a novel bilevel optimization framework for robust point cloud classification, where the internal optimization can effectively defend the adversarial attacks as denoising and the external optimization can learn the classifiers accordingly. As a demonstration, we further propose an effective and efficient instantiation of our approach, namely, Lattice Point Classifier (LPC), based on structured sparse coding in the permutohedral lattice and 2D convolutional neural networks (CNNs) that integrates both internal and external optimization problems into end-to-end trainable network architectures. We demonstrate state-of-the-art robust point cloud classification performance on ModelNet40 and ScanNet under seven different attackers. Our demo code is available at: https://github.com/Zhang-VISLab .

Machine learning (ML) offers several supervised learning algorithms to build classifiers for developing accurate decision support systems. However, the selection of robust classifier for reliable decision making in healthcare domain is three-fold: accuracy, refraining for low confidence decisions, and the cost of decisions. In the field of medical science, there are costs associated with the incorrect and the refraining from decisions, which can have negative implications in devising adequate therapeutic interventions. For example, it may be life threating if a cancer patient is declared as healthy one (misclassification cost) or decision remains pending for some time (rejection cost). In this work we proposed the concept of Accuracy Rejection Normalized-Cost Curves (ARNCCs), which is an extension of Accuracy Rejection Curves (ARCs); a three-dimensional visualization technique to demonstrate the strengths and weaknesses of classification algorithms over different rejection regions and normalized-cost (NC) to select the robust classifier. ARNCCs method holds ARCs plot on two dimensions, in addition it computes NC at third dimension against ratio of false positive costs and ratio of rejection costs obtained at different rejection rates. The proposed three-dimensional graphs have the potential to answer a variety of questions regarding accuracy, rejection rate and NC of a classifier. Six publicly available cancer datasets (four breast cancer and two lung cancer) having clinical parameters obtained from ML repository of University of California, Irvine (UCI) were used to assess the performance of proposed ARNCCs in this study. Empirical results show that ARNCCs provide broad range of decisions to choose the desired parameters (accuracy, rejection rate and NC) for further necessary actions as compared to traditional ARCs method. ARNCCs framework has the ability to more logically compare the performances of classification algorithms in terms of accuracy, rejection rate and NC based scenarios.

The automatic sound event classification (SEC) has attracted a growing attention in recent years. Feature extraction is a critical factor in SEC system, and the deep neural network (DNN) algorithms have achieved the state-of-the-art performance for SEC. The extreme learning machine-based auto-encoder (ELM-AE) is a new deep learning algorithm, which has both an excellent representation performance and very fast training procedure. However, ELM-AE suffers from the problem of unstability. In this work, a bilinear multi-column ELM-AE (B-MC-ELM-AE) algorithm is proposed to improve the robustness, stability, and feature representation of the original ELM-AE, which is then applied to learn feature representation of sound signals. Moreover, a B-MC-ELM-AE and two-stage ensemble learning (TsEL)-based feature learning and classification framework is then developed to perform the robust and effective SEC. The experimental results on the Real World Computing Partnership Sound Scene Database show that the proposed SEC framework outperforms the state-of-the-art DNN algorithm.

In this paper, we introduce a new framework for robust multiple signal classification (MUSIC). The proposed framework, called robust measure-transformed (MT) MUSIC, is based on applying a transform to the probability distribution of the received signals, i.e., transformation of the probability measure defined on their observation space. In robust MT-MUSIC, the sample covariance is replaced by the empirical MT-covariance. By judicious choice of the transform we show that: (1) the resulting empirical MT-covariance is B-robust, with bounded influence function that takes negligible values for large norm outliers, and (2) under the assumption of spherical compound Gaussian noise, the noise subspace can be determined from the eigendecomposition of the MT-covariance. The proposed approach is illustrated for direction-of-arrival (DOA) estimation in a simulation example that shows its advantages as compared to other robust MUSIC generalizations.

Accurate automated classification of brain tumors from magnetic resonance imaging (MRI) is essential for early diagnosis and treatment. This study presents a hybrid framework combining Convolutional Neural Network (CNN) deep features, Large Margin Nearest Neighbor (LMNN) metric learning, and swarm-intelligence optimization for robust four-class classification. Five pretrained CNNs—DenseNet201, MobileNetV2, ResNet50, ResNet101, and InceptionV3—were evaluated on a dataset of 7,023 images categorized as glioma, meningioma, pituitary, healthy. Among these, DenseNet201 provided the highest baseline performance with 92.66% accuracy. LMNN improved feature separability, while Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO) selected compact subsets. The selected features were classified using k-Nearest Neighbor (KNN), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Random Forest (RF). The DenseNet201–LMNN–GWO–KNN configuration, termed DenseWolf-K, achieved the best performance with 99.64% accuracy, establishing it as the optimal implementation of the framework. Robustness and generalizability were further confirmed using an independent external dataset. Model explainability was ensured through feature-level ranking of GWO-selected features and occlusion sensitivity maps, an Explainable Artifical Intelligence (XAI) method. Overall, the proposed DenseWolf-K framework delivers high accuracy, low false-negative rates, compact representation, and enhanced interpretability, representing a reliable and efficient solution for MRI-based brain tumor classification.

Identification of high-risk genes and classification of acute myocardial infarction patients utilizing deep learning in a restricted cohort.

This paper presents a new computational framework for automatic foot classification from digital plantar pressure images. It classifies the foot as left or right and simultaneously calculates two well-known footprint indices: the Cavanagh's arch index (AI) and the modified AI. The accuracy of the framework was evaluated using a set of plantar pressure images from two common pedobarographic devices. The results were outstanding, as all feet under analysis were correctly classified as left or right and no significant differences were observed between the footprint indices calculated using the computational solution and the traditional manual method. The robustness of the proposed framework to arbitrary foot orientations and to the acquisition device was also tested and confirmed.

Pluripotent stem cells can potentially be used in clinical applications as a model for studying disease progress. This tracking of disease-causing events in cells requires constant assessment of the quality of stem cells. Existing approaches are inadequate for robust and automated differentiation of stem cell colonies. In this study, we developed a new model of vector–based convolutional neural network (V-CNN) with respect to extracted features of the induced pluripotent stem cell (iPSC) colony for distinguishing colony characteristics. A transfer function from the feature vectors to the virtual image was generated at the front of the CNN in order for classification of feature vectors of healthy and unhealthy colonies. The robustness of the proposed V-CNN model in distinguishing colonies was compared with that of the competitive support vector machine (SVM) classifier based on morphological, textural, and combined features. Additionally, five-fold cross-validation was used to investigate the performance of the V-CNN model. The precision, recall, and F-measure values of the V-CNN model were comparatively higher than those of the SVM classifier, with a range of 87–93%, indicating fewer false positives and false negative rates. Furthermore, for determining the quality of colonies, the V-CNN model showed higher accuracy values based on morphological (95.5%), textural (91.0%), and combined (93.2%) features than those estimated with the SVM classifier (86.7, 83.3, and 83.4%, respectively). Similarly, the accuracy of the feature sets using five-fold cross-validation was above 90% for the V-CNN model, whereas that yielded by the SVM model was in the range of 75–77%. We thus concluded that the proposed V-CNN model outperforms the conventional SVM classifier, which strongly suggests that it as a reliable framework for robust colony classification of iPSCs. It can also serve as a cost-effective quality recognition tool during culture and other experimental procedures.

CCSA-RF enhanced lightweight multi-scale CNN framework for robust lung cancer classification

Protein classification is a key challenge in bioinformatics, essential for understanding biological functions, molecular interactions, and drug discovery. Yet, the functional landscape of proteins cannot always be inferred from sequence or fold similarity, as structurally divergent proteins may share common biological roles. Since most molecular interactions occur at the protein surface, capturing the geometric and biochemical features of these surfaces is essential for accurate functional inference. In this work, we introduce ProCLR, a graph-based framework that models protein surfaces as simplified, attribute-rich 3D geometric graphs, and applies supervised contrastive learning to learn discriminative embeddings. Our method integrates a geodesicaware clustering strategy to reduce surface complexity while preserving functional features. We evaluate ProCLR using two encoder architectures-Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs)-on the SHREC'21 and SHREC'19 benchmarks. Experimental results show that ProCLR effectively captures functional patterns, achieving competitive performance with significantly reduced graph sizes, as measured by sensitivity, specificity, precision, accuracy, and F1-score.

Accurate classification of fungal morphology is essential for clinical diagnosis. Centralized deep learning approaches raise data privacy concerns and struggle with heterogeneous data distributions. To address these challenges, this study proposes an attention-guided federated learning framework for multi-class fungal morphology classification using the DeFungi dataset. The proposed method integrates enhanced FedAtt attention mechanism at the client level with an attention-aware federated aggregation strategy to improve global model consistency under non-IID data settings. Extensive experiments demonstrate that the proposed framework achieves stable convergence, balanced class-wise performance, and statistically reliable improvements. The proposed model achieved 93.73% accuracy, and the FAA aggregation mechanism improved performance by 8.92% over FedAvg in a privacy-preserving federated learning setting. These results highlight the effectiveness of attention-aligned aggregation for robust and privacy-preserving medical image classification, making the proposed framework suitable for real-world clinical deployment.

Accurate surface defect classification is essential for industrial quality inspection but remains challenging due to the scarcity of annotated samples, high annotation costs, and significant intraclass variations. To tackle these issues, we propose a novel few-shot learning framework for steel strip surface defect classification. Our approach integrates three key components: a self-correlation enhanced (SCE) module for extracting discriminative local region features, a feature refinement module (FRM) for filtering background noise, and a similarity measurement module (SMM) for robust prototype-based classification. In addition, we design a hybrid loss function that jointly optimizes supervised and metric-based objectives to enhance class separability. Experiments on the FSC-20 dataset show that the proposed method outperforms existing approaches in both five-way one-shot and five-way five-shot settings, while also demonstrating strong cross-domain generalization.

Alarm systems are essential safeguards in process industries, providing the first layer of protection by alerting operators to process abnormalities and potential hazards. However, the exponential increase in alarms, mainly driven by their computerized configuration and insufficient rationalization, has compromised their effectiveness. In particular, nuisance alarms distract operators with redundant or misleading information, while alarm floods overwhelm them during process disturbances, increasing the risk of missing critical alarms. Motivated by these challenges, this thesis develops a set of advanced data-driven methods for nuisance alarm reduction, root cause diagnosis, and alarm flood classification and monitoring, along with industrial integration for real-time application. First, a systematic framework is proposed to detect and visualize correlated alarm patterns as enriched networks, where directionality and occurrence order are captured from alarm & event (A&E) data. This representation highlights critical alarms, reveals process interactions, and offers compact and interpretable tools for reducing nuisance alarms. Second, a data-driven technique is introduced to construct adjacency matrices of control loops from alarm correlations. This approach enables preliminary root-cause diagnosis of plant-wide oscillations without relying on detailed process models, while significantly reducing computational complexity in large-scale facilities. Third, three complementary frameworks are developed for alarm flood analysis and monitoring: (i) a visual analytics-based convolutional matching framework for interpretable alarm flood classification (ii) a deterministic hashing approach, Combinatorial Temporal Fingerprint Hashing (CTFH), for robust and scalable classification, and (iii) a context-enriched embedding model capturing temporal, sequential, and priority information of alarms to achieve adaptive, interpretable real-time flood classification. Finally, an existing smart data analytics framework is enhanced with features for secure SQL and OPC UA-based data extraction from client databases and automated reporting, aligning the platform with Industry 4.0 standards for connectivity, agility, and intelligence. The effectiveness of the proposed methods is demonstrated through case studies on benchmark processes and industrial datasets. Collectively, the contributions provide a comprehensive framework for the next generation of alarm systems, advancing robustness, scalability, and interpretability, while offering practical decision-support tools for operators in modern process industries.

The steady-state visual evoked potential (SSVEP), a non-invasive EEG modality, is a prominent approach for brain–computer interfaces (BCIs) due to its high signal-to-noise ratio and minimal user training. However, its practical utility is often hampered by susceptibility to noise, artifacts, and concurrent brain activities, complicating signal decoding. To address this, we propose a novel hybrid deep learning model that integrates a multi-channel restricted Boltzmann machine (RBM) with a convolutional neural network (CNN). The framework comprises two main modules: a feature extraction module and a classification module. The former employs a multi-channel RBM to unsupervisedly learn latent feature representations from multi-channel EEG data, effectively capturing inter-channel correlations to enhance feature discriminability. The latter leverages convolutional operations to further extract spatiotemporal features, constructing a deep discriminative model for the automatic recognition of SSVEP signals. Comprehensive evaluations on multiple public datasets demonstrate that our proposed method achieves competitive performance compared to various benchmarks, particularly exhibiting superior effectiveness and robustness in short-time window scenarios.