FallbackWalk: A Random Walk Based Fallback for Heterogeneous Information Network

Graph Embedding, as a network representation learning technology, can analyze information networks and represent nodes as low dimensional dense vectors, which shows good performance and attracts a lot of research interests. Heterogeneous information networks contain rich information. However, the current models still have some limitations in capturing semantic information in information networks. Most models mainly focus on how to aggregate the semantic information in the information network but do not consider the differences of nodes in the feature extraction process. In order to solve this problem, a graph embedding model based on fallback strategy (FallbackWalk) is proposed. FallbackWalk takes into account the differences of nodes in the information network, and make more use of neighborhood nodes by random walk strategy based on fallback, and the Skip-gram model is used to train and get the vector representation of nodes. To validate our proposed network representation learning model, we present experiments on two real-world datasets and two representation learning tasks: node classification and node clustering. The node representation of FallbackWalk can capture the node characteristics more accurately. The experimental results show that the node representation of FallbackWalk outperforms state-of-the-art models on different downstream tasks.