CRWMS: Bipartite Network Embedding based on Constrained Random Walk and Mixed-Skipgram

A bipartite network is a basic representation model in recommendation systems, in which the explicit links (e.g., representing user-item rating information) only exist between heterogeneous nodes (e.g., users and items). Most traditional bipartite network embedding methods only consider the explicit network structure, but ignore semantic relations and rating information therein. To solve this deficiency, we designed a new bipartite network embedding approach based on the Constrained Random Walk and Mixed-Skipgram (CRWMS). Specifically, we propose a constrained random walk strategy that produces the mixed local sampling sequences, and further adopt the splitting operation to preserve both explicit and implicit structural information. Furthermore, we propose a novel and effective mixed-skipgram model to discern node representation vectors through joint training of explicit and implicit structural sequences. We compare CRWMS with five state-of-the-art network embedding methods on three real-world bipartite networks. Experiments show that our approach improves performance in edge classification, edge clustering, recommendation and link reconstruction tasks.