Semantic neighborhood-aware fuzzy clustering for balanced text categorization.

Abstract

Text categorization remains a challenging task due to the inherent ambiguity of natural language, class overlap, and imbalanced topic distributions. Traditional Fuzzy C-Means (FCM) clustering, although widely used for soft text classification, is highly sensitive to initialization and tends to favor dense or majority clusters, leading to suboptimal performance on complex textual datasets. This study proposes a Bias-Adjusted Spatially Constrained Fuzzy C-Means (BASC-FCM) algorithm to address these limitations. The proposed approach extends the classical FCM objective function by integrating a spatial constraint that enforces semantic neighborhood consistency in the feature space and a bias adjusted term that mitigates dominance by large clusters. Spatial proximity is interpreted as semantic similarity, encouraging documents with similar lexical structure to share comparable membership distributions, while the bias term reweights cluster influence to improve minority class representation. Experiments were conducted on a real-world news dataset comprising 7600 English articles across four categories. Performance was evaluated using accuracy, precision, recall, F1-score, and computational time. Results demonstrate that BASC-FCM significantly outperforms standard FCM and its constrained variants, achieving improvements of up to 63.6% in accuracy and 81.0% in F1-score, while maintaining acceptable computational overhead. The findings indicate that jointly incorporating spatial constraint and bias adjustment yields an enhanced and balanced clustering framework for text categorization tasks.