Learning to extract and aggregate contexts for link prediction in heterogeneous graphs

Abstract

Many diverse real-world graph datasets are heterogeneous graphs, and link prediction on these graphs is a fundamental task. The current trends of link prediction on heterogeneous graphs emphasize leveraging contextual information from either a path between a source node and a target node, or a sub-graph sampled around these two nodes. However, these approaches face limitations in identifying only beneficial contextual nodes around source and target and then effectively aggregating the representations of these nodes for improving overall prediction accuracy. To address these limitations, we claim that carefully-extracted context nodes can aid in accurate link prediction, and these context nodes should be similar to a source node or a target node in a representation space. To this end, we propose a new link prediction framework LEACH which learns to extract the beneficial context nodes and to aggregate their representations in heterogeneous graphs. Specifically, our approach involves three steps to learn: (i) generating heterogeneity-aware representations of nodes in the heterogeneous graph, (ii) selecting the context nodes based on the relatedness to the source and target nodes; and (iii) aggregating the representations of the context nodes to obtain the source and target representations. Extensive experiments demonstrate that LEACH significantly outperforms existing baselines on three publicly available heterogeneous graph datasets. We provide analytical insights into the rationale behind the superior performance of LEACH on link prediction.