This study investigates the integration of active learning and self-labeling in deep image classification under constrained annotation budgets. The objective was to design a hybrid semi-supervised learning strategy that combines expert-guided sample selection (active learning) with automatic label assignment (self-labeling) while introducing a correction mechanism to mitigate the effects of erroneous pseudo-labels. The methodology was validated experimentally using the CIFAR-10 dataset and a ResNet-18 model. Experiments were conducted with annotation budgets of 5
Machine learning has become a core component of modern credit scoring systems, but it often exacerbates gender disparities due to structural biases in the data. These biases primarily arise from two sources: the underrepresentation of certain groups and confounding correlations between sensitive attributes and predictive features. Existing mitigation techniques typically focus on only one of these sources, resulting in limited fairness improvements or notable accuracy degradation. To address this problem, this paper proposes a cascaded bias mitigation method that targets different types of bias in each stage, because the two sources operate through different mechanisms. Pre-processing stage adopts a data-level rebalancing operator to mitigate data imbalance, and in- and post-processing stages jointly address structural bias induced by sensitive-confounded feature-label correlations. The whole pipeline is optimized using a multi-objective Bayesian optimization that efficiently explores combinations of mitigation techniques and hyperparameters to balance fairness and predictive performance. Evaluated on three real-world credit scoring datasets, the proposed method reduces gender bias by up to 99
This paper introduces an incremental method to make clustering models more interpretable by adding explanatory variables. Using origin–destination frequency tensors and HDBSCAN for unsupervised clustering, the approach incrementally adds exogenous variables and assesses their effects while keeping cluster structure intact. This process clarifies mobility pattern analysis and offers a framework for enhancing interpretability in other models.
Variational Quantum Classifiers (VQCs) are hybrid algorithms that combine classical machine learning with quantum computing techniques, aiming to improve upon traditional classification approaches. In this paper, we test the hypothesis that daily price variations in selected Romanian energy-sector stocks contain predictive information about the daily direction of price variation in a major Romanian pharmaceutical stock. We further examine whether a shallow variational quantum classifier can model these cross-sector nonlinear dependencies more effectively than standard classical classifiers. Our expectation is not a provable computational quantum advantage, but a possible empirical modeling benefit arising from quantum feature embedding and entangling layers in a low-dimensional classification setting. The results demonstrate that VQC2’s quantum feature mapping effectively captures nonlinear interactions that classical models would overlook, increasing its predictive potential.
The accelerating proliferation of the Internet of Things has increased the exposure of connected devices to cyber threats, underscoring the need for research aimed at advancing intrusion detection systems. Representing network traffic as images enables the use of convolutional neural networks for intrusion detection. This work adopts a feature imaging approach based on mutual information and applies it to the NF-UNSW-NB15-v3 dataset to create structured inputs for supervised machine learning. A lightweight convolutional architecture subsequently processes the transformed feature images to classify traffic as benign or as a specific attack type. Experimental results show a strong discrepancy between overall accuracy and macro-F1: although the model attains high accuracy, its multiclass performance on minority attack categories remains limited under the severe class imbalance of NF-UNSW-NB15-v3, in addition to requiring higher computational cost and longer training times.
