Multi-scale representation learning for heterogeneous networks via Hawkes point processes

In the field of dynamic heterogeneous network representation learning, current research methods have certain limitations. These limitations are mainly observed in the manual design of meta-paths, the handling of node attribute sparsity, and the fusion of dynamic heterogeneous information. To overcome these challenges, this paper presents a multi-scale representation learning method for heterogeneous networks via Hawkes point processes called MSRL. MSRL models the self-excitation effect among historical events by integrating the Hawkes process and captures the facilitating effect of external structures on event occurrence through a ternary closure process. This study employs the integration of time series analysis with neighbourhood interaction information to automate the extraction of the node pair representation. The MSRL model treats edges as time-stamped events, which not only captures the temporal dependencies between events, but also addresses the imbalance between different node types and the challenge of information fusion from a multi-granularity perspective. In particular, the model enhances the accurate estimation of the probability of node pairs forming edges by analysing the interactions between node pairs and their neighbours, which significantly improves the accuracy of tasks such as prediction. To validate the effectiveness of the MSRL model, an extensive experimental evaluation is conducted in this paper. The experimental results show that the MSRL model outperforms existing baseline models on several benchmark datasets, demonstrating its significant advantages and potential applications in the field of dynamic heterogeneous network representation learning.