International Journal of Information Systems and Computer Technologies

A Skeleton-Based Geometric Learning Framework for Compact and Explainable Supervised Classification

2026-07-18T16:27:27+01:00

Efficient supervised classification is challenging for high-dimensional and large-scale data because typical classifiers have large storage requirements, extensive parameter optimization, and limited geometric interpretability. This study introduces the Linear Skeleton Classifier (LSC), a geometric supervised learning model where each class is defined by a minimalist piecewise-linear skeleton built from a set of per-class k-means centroids. A query sample is classified into the class with minimum point-to-line segment distances, which offers a reduced storage complexity of O(C×k), where C is the number of classes and k is the number of centroids per class. The effectiveness of the approach is demonstrated with classification experiments conducted on seven standard benchmark datasets using a 70/30 train-test splitting approach. LSC achieves test accuracies of 93.33% (Iris), 96.30% (Wine), 79.37% (Sonar), 95.91% (Breast Cancer), 90.65% (Dry Bean), 82.73% (MNIST_784 subset), and 79.01% (Heart Disease). A comparative study has also been carried out using standard classifiers such as k-NN, SVM (linear/RBF), Naive Bayes, and Decision trees, which illustrate that LSC achieves better trade-offs between performance, a small reduction in accuracy, robustness against noise, and insensitivity to bias-variance issues. Moreover, bias-variance analyses demonstrate that among LSC and standard classifiers, LSC provides a near-optimal balance between minimum error and bounded bias/variance. LSC shows favorable results not just in predictions but also eases intelligibility in skeleton segmentation through path profiling and provides geometric clues regarding how class samples progress and how outliers can be weeded out. Thus, LSC a scalable, explainable alternative for manifold classification in terms of compression, predictability, and diagnostic power.

An Efficient Vision Transformer-Based Deep Hashing Framework for Semantic Image Retrieval

2026-07-18T16:38:23+01:00

X‑HashNet presents a transformer-based supervised hashing scheme to enable highly efficient fashion image retrieval. Leveraging the architecture of the popular vision model DeiT-Small, the model replaces convolutional architectures with a global attention-based representation paradigm. Evaluated on the Fashion Minority dataset (Fashion-Mnist), which consists of 70,000 grayscale images from 10 apparel categories, X-HashNet is used to make 64-bit binary embeddings from the raw inputs that are optimized for retrieval based on the same Hamming distance. The pipeline combines 5 key stages: ViT-adaptable patch embedding, transformer-based feature extraction, supervised bottleneck hashing, multi-objective optimization, and FAISS-based binarization and indexing. The model has a mean Average Precision (mAP@100) of 0.9348, which is a new state-of-the-art benchmark for hashing on Fashion-MNIST. Diagnostic analyses confirm the best utilization of codes: the average bit activation is 0.4907; inter- and intra-class Hamming distances are 2.29 & 0.28 bits, respectively; and hash stability is 84.20 percent. The bit redundancy score of 0.2775 and near-ideal entropy distribution mean efficient encoding of information in all the hash dimensions. Empirical results also validate a strong level of generalization, yielding Precision@1 of 93.39% and maintaining stable performance at deeper response latencies (P@5-P@100 ~93%). From a systems perspective, the average query time of 0.1738 milliseconds and throughput of more than 5 754 queries per second make X HashNet suitable for large-scale deployment, where a mere 8 bytes per image are used to index the images. Visual attention maps validate the model's ability to both localize and maintain important structural features (e.g., silhouettes and textures of clothing). Collectively, these results show evidence that transformer-based hashing not only outperforms CNN counterparts in retrieval accuracy, but also provides a scalable, industrially viable foundation for real-time search and recommendation systems for fashion.

Adaptive Fuzziness Weighting for Dynamic Instance Replication and Improved Classifier Generalization

2026-07-18T16:17:04+01:00

The performance ceiling is often how current classifiers deal with ambiguous instances that are very close to the decision boundary in the training set. Conventional instance selection and weighting methods often rely on error-based measures or predefined rules, thereby neglecting the important relationship between an instance's informativeness to the model and the instance's own uncertainty (fuzziness). The Adaptive Fuzziness Weighting (AFW) strategy is introduced in this paper for a new three-phase meta-learning system in which dynamic data augmentation and robust boundary refinement are developed. The base classifier (NNRw: Neural Network with Random Weights) is used to calculate the fuzziness score (F) of an instance, based on the margin of certainty of the posterior probabilities. Importantly, during the second phase, this fuzziness score is automatically converted into an adaptive replication weight (????) by a lightweight Ridge Regression meta-learner, which efficiently learns the best influence factor from each instance, depending on its fuzziness and initial correctness. The last step performs a fuzziness-regulated replication of the instances in the training set, as a means of enhancing it, by preferentially multiplying high-confidence instances. Using 10-fold cross-validation on ten UCI benchmark datasets, the AFW approach yielded significant improvements. The results showed that mean classification accuracy increased from 0.7134 to 0.9266, a 21.32 percentage-point improvement on average. The most prominent improvement was observed in datasets that were naturally complex (e.g., Wine and Automobile), demonstrating AFW's ability to stabilize and fine-tune decision boundaries in challenging feature spaces. AFW is a significant improvement on data augmentation because it offers a generalized, analytically motivated approach to assigning an objective measure of uncertainty to an instance, and then using this uncertainty to drive a training priority.

An Explainable Clinical Decision-Support Framework for Brain Tumor Detection and Classification from MRI

2026-08-14T19:54:30+01:00

Brain tumors are among the most life-threatening neurological disorders, and early, accurate diagnosis is critical for improving treatment outcomes and patient survival. MRI is the gold-standard modality for brain tumor visualization. However, proper MRI interpretation requires specialized radiological expertise, which may be limited in resource-constrained healthcare settings such as Pakistan. Furthermore, traditional deep learning methods have relied on ImageNet pre-trained models, which may not capture MRI-specific representations and usually lack a clinic-deployable decision-support system. To address these challenges, this study introduces an Explainable Clinical Decision-Support Framework for automatic brain tumor detection and four-class classification of glioma, meningioma, pituitary tumor, and no-tumor from MRI. Gradient-weighted Class Activation Mapping (Grad-CAM) generates per-prediction heatmap overlays that highlight the area contributing to each classification decision, enhancing interpretability. A two-stage learning-rate refinement strategy was employed to improve convergence and classification performance. The framework was trained and evaluated using a stratified 70%/15%/15 % split for the training, validation, and test sets, respectively, and achieved 94.79% accuracy. Comparative performance analysis against a MobileNetV2 baseline shows that the proposed customized CNN model performs better in terms of accuracy for the two difficult-to-diagnose classes of glioma and meningioma. Results indicate that the developed framework has the potential to give an accurate and interpretable classification of brain tumors from MRI imaging. Nevertheless, external validation on diverse multi-center and multi-national datasets, along with prospective clinical studies, is needed.

A Layer-Specific Taxonomy of Security Challenges in Internet of Things

2026-08-21T16:56:29+01:00

IoT technology has allowed numerous devices, networks, and services to communicate with each other on an extensive scale. With increasing connectivity, IoT security has become a major concern and expands IoT attack surfaces in general. Because of the diversity of IoT system design, limited IoT devices’ resources, and extensive information exchange between devices, IoT systems are vulnerable to complex threats. However, most previous studies present the threats of IoT in a dispersed form. This study contributes to knowledge by proposing a taxonomy of layer-specific security challenges in IoT by combining findings from previous studies and experts’ evaluations. An exhaustive list of challenges was gathered through the comprehensive literature review. A five-point Likert-scale questionnaire was conducted with 70 participants with at least six years of relevant experience. PCA was deployed to reduce and extract the crucial challenges. 12 challenges emerged as crucial challenges. The extracted challenges have been categorized into 4 dimensions to devise a taxonomy. The proposed taxonomy offers a structured view to identify important security challenges in different IoT layers. It also facilitates experts in prioritizing security challenges and selecting suitable countermeasures.