Granularity Gating-Guided Label-Specific Metric Learning for Multi-Label Classification

Label-specific local metric learning improves nearest-neighbor multi-label classification by adding label- and region-specific private transformations to a shared one. The regions assigned to a training anchor, its support, select the anchor's transformation and also determine which private transformations its training constraints update. Existing methods partition the data, forcing instances with overlapping semantics into one region. Soft cluster memberships avoid this choice but send every anchor's supervision to every region, and with uniform memberships all private transformations of a label become identical. We therefore restrict each local metric to the anchors that provide enough evidence for it. Granularity-Gated Fuzzy Metric Learning (GFML) sets this restriction with a gate derived from the principle of justifiable granularity. Each region receives a granule whose radius balances rival-adjusted membership coverage against specificity. An anchor inside a granule uses and trains only the private transformations whose granules contain it, so excluded transformations receive none of its supervision. On eight benchmarks with two KNN classifiers, GFML matches or improves on the strongest metric-learning and specialized baselines; ablations show that the gate improves on dense support.