New Model Improves Multi-Label Node Classification on Graphs

Graphs are ubiquitous data structures, and nodes often possess multiple labels, such as proteins with various functions or users with diverse interests. Accurately classifying these multi-label nodes (MLNC) is a critical challenge. While existing methods leverage graph neural networks (GNNs) for label co-occurrence or incorporate label embeddings, they frequently overlook the complex, dynamic influences labels exert on each other within non-Euclidean graph data. To address this gap, researchers propose Label Influence Propagation (LIP). This model first decomposes the message passing process in GNNs into propagation and transformation operations. It then meticulously analyzes and quantifies the influence correlations between labels within each operation. Based on these insights, LIP constructs a dedicated label influence graph. Through this influence graph, LIP propagates high-order influences, dynamically adjusting the learning process. It amplifies labels with positive contributions while mitigating those with negative influence, leading to more accurate classifications. Comprehensive evaluations on benchmark datasets demonstrate that LIP consistently surpasses state-of-the-art methods across various settings, proving its effectiveness for MLNC tasks.