Mutual Optimization of Label Prediction and Node Representations for Multi-Label Node Classification

Abstract

Multi-label node classification requires capturing complex label-co-occurrence patterns, which are critical for accurate predictions. However, traditional methods often treat label prediction as an independent task, overlooking the semantic relationships and dependencies between labels. This limitation hinders the ability to model intricate label-co-occurrence information. To address this issue, we propose the Mutual Optimization of Label-Prediction and Node-Representations (MOLN) framework. MOLN integrates two interrelated components: Semantic-Aware Contrastive Learning (SACL) and the Label-Interaction-Enhancement Network (LIE). SACL operates in the node-representation space, aligning embeddings with label-co-occurrence patterns to capture meaningful semantic relationships. LIE, on the other hand, works in the label-prediction space, explicitly modeling mutual dependencies among labels to refine predictions. The natural synergy between these components ensures mutual optimization: SACL enriches node features for prediction, while LIE improves predictions to guide better semantic alignment in SACL. By bridging the representation and prediction spaces, MOLN effectively addresses the label-co-occurrence problem, achieving state-of-the-art performance on benchmark datasets with significant improvements.