Multi-angle information aggregation for inductive temporal graph embedding

Abstract

Graph embedding has gained significant popularity due to its ability to represent large-scale graph data by mapping nodes to a low-dimensional space. However, most of the existing research in this field has focused on transductive learning, where fixed node embeddings are generated by training the entire graph. This approach is not well-suited for temporal graphs that undergo continuous changes with the addition of new nodes and interactions. To address this limitation, we propose an inductive temporal graph embedding method called MIAN (Multi-angle Information Aggregation Network). The key focus of MIAN is to design an aggregation function that combines multi-angle information for generating node embeddings. Specifically, we divide the information into different angles, including neighborhood, temporal, and environment. Each angle of information is modeled and mined independently, and then fed into an improved gated recuttent unit (GRU) module to effectively combine them. To assess the performance of MIAN, we conduct extensive experiments on various real-world datasets and compare its results with several state-of-the-art baseline methods across diverse tasks. The experimental findings demonstrate that MIAN outperforms these methods.

Keywords: Graph embedding; Inductive learning; Multi-angle information; Temporal graph.

Figure 1. Overall framework of MIAN.

This framework consists of two steps: the first step is to model information from different angles and generate information embeddings separately, and also initialize node embeddings for later updates. The second step is to aggregate these information embeddings using an improved GRU module to generate the final (new) node embeddings.

Figure 2. Parameter sensitivity study with different node activate threshold values.

Figure 3. Ablation study on different variants of the proposed MIAN method.

Figure 4. Convergence analysis on all datasets with loss value evolution.