Pattern Analysis and Applications Computer Vision and Pattern Recognition
Local-structure-aware fuzzy matrix self-enhancement for reliable clustering ensembles in pattern analysis
Abstract
In clustering-based pattern analysis, the reliability of inter-sample similarities affects the quality of consensus partitions. Fuzzy Matrix Self-Enhancement (FMSE) refines these similarities using global fuzzy representativeness, but its sample weights do not explicitly describe local neighborhood contrast. We introduce Local Confidence of Instance (LCI), which measures the normalized gap between a sample’s top-t reconstructed similarities and the remaining off-diagonal similarities. LCI complements the global Fuzzy Representativeness Index through geometric fusion and sample reweighting of the reconstructed fuzzy co-association matrix. Evaluation on seven University of California, Irvine benchmark datasets shows a mean clustering accuracy (ACC) increase of 0.0067, or 0.67 percentage points, over FMSE-a, with ACC gains on five of the seven datasets. Full ablations, controlled synthetic experiments, fuzzy-membership perturbations, and parameter analyses examine the contributions and limitations of the weighting scheme. The two-sided Wilcoxon ACC test yields p = 0.0781, so the average gain does not establish statistical significance at the 0.05 level. The results support a data-dependent enhancement of FMSE; LCI supplies local similarity contrast rather than a universal instance-reliability or outlier-detection guarantee.