Meta structure over heterogeneous information network embedding

A heterogeneous information network(HIN) is a type of graph that contains various types of nodes or edges. The purpose of representing a HIN is to use low-dimensional, dense vectors to capture high-dimensional graph data. Many practical applications can be modeled as heterogeneous networks. For example, in an e-commerce network, merchants of different brands and users can be represented as different types of nodes, while relationships such as purchasing and following can be represented as different types of edges. Representation learning on such networks is beneficial for the construction of recommendation systems. Currently, the main representation learning methods for heterogeneous networks are the random walk shallow models and neural network deep models. However, since the meta-path in these methods cannot represent complex structures, this paper focuses on a meta-structure-based heterogeneous network representation learning method. It controls the random walk using the meta-structure, obtains the sampled node sequence, and uses it as the input for the word embedding training model, finally obtaining the node embeddings. In this paper, the meta-structure is used to guide the random walk. According to experimental comparisons, this approach achieves a better representation of heterogeneous networks.