Extending Heterogeneous Hypergraph Embedding LBSN2vec Toward Behavioral Analysis of Online Product Reviewing

Abstract

From the point of view of representation learning, analysis and mining of complex network data ordinarily require an effective feature vector respresentation for the nodes. Several network embedding methods such as node2vec were developed for classical graphs, where an edge connects two homogeneous nodes. However, many of social media datasets such as LBSNs (Location-Based Social Networks) and online product reviews can be naturally modeled as heterogeneous hypergraphs, where an edge connects more than two heterogeneous nodes. Recently, LBSN2vec, a heterogeneous hypergraph emebedding method, has been presented for LBSN datasets with classical edges (friendships) connecting two user-nodes, and it has been empirically shown that LBSN2vec outperforms state-of-the-art graph embedding techniques including node2vec for both friendship and location prediction tasks. In this paper, we focus on a heterogeneous hypergraph constructed by an online product review dataset in which there is no classical graph structure among the user-nodes but there are review documents posted by the users, and we attempt to extend LBSN2vec to such heterogeneous hypergraph data. Using Rakuten Ichiba dataset, we empirically demonstrate the effectiveness of the proposed extension method in terms of behavioral prediction.