Information granularity provides a framework for machine intelligence to simulate human cognitive processes in problem-solving, enabling machines to make more flexible and adaptive decisions in complex data environments. However, most existing information granularities are based on fuzzy c-mean (FCM) or K-means clustering for granularity construction. This kind of convex data structure clustering has difficulty in recognizing some irregular nonconvex data types, and can only be filled irregularly by increasing the cluster classes, which is easy to lose the wholeness and accuracy of the data. The density-based spatial clustering of applications with noise (DBSCAN), a classical density-based spatial clustering method, is well-suited for both convex and nonconvex datasets. Building on the versatility of DBSCAN, a new density-based clustering method called stacked density granulation (SDG) is proposed, which constructs information granularities capable of effectively describing both convex and nonconvex data. This approach addresses the limitations of traditional information granularity. By applying the newly constructed information granularities to the feature space, the importance of features can be measured. To enhance this process, a heuristic feature selection method called density granular feature selection (DGFS) is introduced. DGFS constructs a low-dimensional feature space by aggregating and discretizing the information granularities, retaining only the most relevant features. To demonstrate the superiority and effectiveness of the DGFS algorithm, 12 publicly available datasets are utilized, and its performance is compared with other feature selection methods across four different classifiers. The experimental results and the statistical significance test indicate that DGFS consistently outperforms the competing methods.
Wentao Li, Xuan-Zhen Zhao, W. Pedrycz et al.· IEEE Transactions on Cyberne...· 0 citations
In this study, we propose an incrementally expanding fuzzy neural module network (FNMN) designed to effectively handle both low- and high-dimensional problems without relying on dimensionality reduction techniques. The proposed framework adopts a modular and hierarchical architecture, in which univariate fuzzy neural modules (UFNMs) are incrementally selected and connected according to their representational capability, quantified by the coefficient of determination. The variable-specific univariate fuzzy rule architecture avoids exponential rule growth while preserving interpretability. A residual-driven hierarchy incrementally selects informative modules and progressively refines the model. A hybrid learning strategy combines efficient modulewise learning based on least squares error estimation with global fine-tuning via backpropagation (BP). Adam-based optimization is adopted to enhance convergence stability and reduce sensitivity to learning-rate settings. Extensive experiments on 28 publicly available benchmark datasets demonstrate the effectiveness of the proposed approach. The proposed method achieves an average performance improvement of 19% compared with a conventional fuzzy clustering-based model across diverse benchmarks. Statistical significance tests further confirm that the proposed model significantly outperforms recent neurofuzzy systems. Notably, competitive predictive performance is attained while model complexity is reduced by more than two orders of magnitude relative to deep learning approaches. These findings highlight the efficiency and scalability of the proposed framework.
Eun-Hu Kim, Hao Huang, Zheng Wang et al.· IEEE Transactions on Cyberne...· 0 citations
In Class-Imbalanced Semi-Supervised Learning (CISSL), classifiers suffer from severe confirmation bias, particularly when the class distribution of unlabeled set are unknown or mismatched with that of the labeled set. Existing debiasing methods often rely on impractical assumptions for classifier bias estimation, such as known class priors or risky external proxies. To overcome these limitations, we propose Intrinsic Logit-Based Debiasing (ILBD), a robust post-hoc framework that estimates bias directly from task-relevant data without external dependencies and class priors. ILBD constructs a comprehensive bias estimator by decomposing output logits into two distinct components: (1) Label-Free Intrinsic Bias, derived from non-target K-1 logits by masking target-class maximum logits to capture the model's background distributional skew; and (2) Label-Guided Learning Bias, derived from target-class maximum logits to quantify confidence disparities between head and tail classes arising from varying learning difficulties. To ensure estimation reliability, we further employ a class-adaptive threshold to filter noisy pseudo-labels. By subtracting this estimated bias at inference, ILBD effectively rectifies the classifier's bias directly from the inherent statistical patterns of the training data, without the need for target-prior assumptions or external data. Extensive experiments on CIFAR-10/100-LT, STL-10-LT, and Small-ImageNet-127 demonstrate the effectiveness of ILBD. Our code is available at https://github.com/aroid721/ILBD.
