iHOMER+: Incremental Multi-View Output-Specialization for Streaming Multi-Label Learning

• Proposes iHOMER+, a multi-view ensemble for drifting multi-label streams • Reviews local (problem transformation) and global (algorithm adaptation) strategies • Introduces output-specialization methods that cluster labels for hybrid modeling • Models correlations, across label, instance, and feature spaces • Combines three concurrent learners for robust drift adaptation Mining multi-label data streams presents significant computational challenges due to exponentially large output spaces and evolving distributions. Most incremental multi-label learners rely on ensembles of lightweight local models that decompose the task into simpler single-label components. While efficient, these methods struggle with extremely sparse labels and complex label dependencies. Global approaches, which consider the full output space, can better exploit label correlations to improve predictive performance. To combine the advantages of both strategies, we introduce iHOMER+, an ensemble of output-specialized models predicting subsets of the output space using global predictors. By partitioning the label space into multiview clusters that balance between fully local and global extremes, iHOMER+ captures both label correlations and sparsity patterns. The method further incorporates drift detection at both global and local levels, enabling dynamic restructuring of label partitions and subtrees, validated through an alternate background replacement process. Experiments on nine synthetic datasets with annotated concept drifts demonstrate that iHOMER+ improves predictive performance by 5% over the current state-of-the-art global method and by 29% over six benchmark local methods. Code and datasets are publicly available.