Network Sampling Using k-hop Random Walks for Heterogeneous Network Embedding

Capturing neighborhood information by generating node sequences or node samples is an important prerequisite step for many of the neural network embedding approaches. Majority of the recent studies on neural network embedding exploit random walk as a sampling method, which traverses through adjacent neighbors to generate the node sequences. Traversing through only immediate neighbor may not be suitable particularly for heterogeneous information networks (HIN) where adjacent nodes tend to belong to different types. Therefore, this paper proposes a random walk based sampling approach (RW-k) which generates the node sequences such that adjacent nodes in the sequence are separated by k edges preserving the k-hop proximity characteristics. We exploit the node sequences generated using RW- k sampling for network embedding using skip-gram model. Thereafter, the performance of network embedding is evaluated on future co-authorship prediction task over three heterogeneous bibliographic networks. We compare the efficacy of network embedding using proposedRW-k sampling with recently proposed network embedding models based on random walks namely, Metapath2vec, Node2vec and VERSE. It is evident that the RW- k yields better quality of embedding and out-performs baselines in majority of the cases.