AAHIN: attribute-aware heterogeneous information network representation learning for data mining

Purpose Heterogeneous graphs, composed of diverse nodes and edges, are prevalent in real-world applications and effectively model complex web-based relational networks, such as social media, e-commerce and knowledge graphs. As a crucial data source in heterogeneous networks, Node attribute information plays a vital role in Web data mining. Analyzing and leveraging node attributes is essential in heterogeneous network representation learning. In this context, this paper aims to propose a novel attribute-aware heterogeneous information network representation learning algorithm, AAHIN, which incorporates two key strategies: an attribute information coverage-aware random walk strategy and a node-influence-based attribute aggregation strategy. Design/methodology/approach First, the transition probability of the next node is determined by comparing the attribute similarity between historical nodes and prewalk nodes in a random walk, and nodes with dissimilar attributes are selected to increase the information coverage of different attributes. Then, the representation is enhanced by aggregating the attribute information of different types of high-order neighbors. Additionally, the neighbor attribute information is aggregated by emphasizing the varying influence of each neighbor node. Findings This paper conducted comprehensive experiments on three real heterogeneous attribute networks, highlighting the superior performance of the AAHIN model over other baseline methods. Originality/value This paper proposes an attribute-aware random walk strategy to enhance attribute coverage and walk randomness, improving the quality of walk sequences. A node-influence-based attribute aggregation method is introduced, aggregating neighboring node attributes while preserving the information from different types of high-order neighbors.