Mining Underlying Structural Insights via High-Order Interactions
Meilin Liu, Wenping Zheng · 2024
Complex networks effectively model data by representing entities as nodes and interactions as edges, revealing intricate association patterns and highlighting inherent structural information. High-order multivariate interactions play a crucial role in characterizing interactions among a group of nodes, providing insights into the complex structures inherent in data. However, multivariate interactions are not always explicitly present, necessitating the utilization of binary interaction information and attribute data to effectively mine them. Message passing allows for the efficient propagation of information through the network, and helps in understanding dynamic processes and discovering of underlying structures and patterns in the data. Designing a message passing mechanism based on multivariate interactions can uncover rich underlying structures in the data, such as clusters and functional modules. We propose a framework that utilizes multivariate interactions for constructing a high-order network and propagating messages within it to uncover the underlying interaction structures in complex data. We present a high-order network construction method using a learning-based approach to identify intricate multivariate interactions from binary interaction information and attribute features. We design a multi-scale message passing paradigm to transmit messages at the node level, edge level and graph level of a high-order network within a unified framework. Finally, we integrate the multi-scale message passing paradigm with a label propagation process to uncover underlying clusters within the data. Experiments conducted on binary network datasets and structured datasets demonstrate that our proposed high-order interactions-based label propagation mechanism outperforms traditional pairwise network algorithms in mining underlying structures.