A Novel Conviction‐Driven Meta‐Learning Framework for Label‐Dependent Multi‐Label Classification

ABSTRACT

Multi‐Label Classification (MLC), a common problem in domains like biology, semantic scene analysis, and text classification, involves association of a sample with more than one class label. Effective utilization of label information helps improve the prediction performance of multi‐label classifiers. Stacked Generalization is a two‐level multi‐label classification technique that models label dependencies by feeding base classifier predictions into the meta classifiers. However, existing techniques primarily utilize symmetric association measures and weight all labels equally, resulting in noise and ignoring directional relationships. This work proposes Conviction‐based Stacking for Multi‐Label Classification (ConvST), a novel framework that uses conviction, an asymmetric, rule‐based association measure to determine a label‐specific and directionally relevant subset of meta‐level inputs. Conviction between class labels is computed based on predictions of base classifiers; thus, enabling dynamic and model‐aware selection of relevant class labels and ignoring weak or redundant dependencies during meta‐learning. This results in compact, discriminative meta‐features that improve the prediction efficiency and generalization performance. The proposed framework is validated on 16 benchmark multi‐label datasets of varying domains, comparing its performance against eight state‐of‐the‐art techniques across 13 performance measures. Statistical tests demonstrate that ConvST achieves consistent and significant improvements against most techniques; thus, confirming its efficacy and the theoretical robustness. This work integrates conviction‐based, directional label dependency modeling into Stacking MLC, and is a novel and generalizable framework for multi‐label classification.