Node Embedding over Attributed Bipartite Networks in Dynamic Context

This work seeks an effective embedding representation by the exploitation of both the structure and attribute aspect of attributed bipartite networks in a dynamic context. Most of the plain or attributed network embedding works are limited to the static type homogeneous networks, but in reality, networks often evolve that certainly effect the node homophilic relations. Hence, we propose DABNE embedding framework that investigates the dynamics of evolved nodes/links over time through targeting both the explicit and implicit node interactions of different and same entity types of a bipartite network. Primarily, it targets the dynamics of the network (structural change) and dynamics on the network (explicit/implicit node interactions either through structure or attributes similarities) on the network. We adopt the KL divergence method to model the node’s explicit interactions, and for implicit node interactions, we employ the Skipgram model. Lastly, we concatenate these independently learned latent embeddings to perform the classification (multi-class and multi-label) and clustering task. Experiments are conducted on two real-world networks with both structure and attributed based embedding methods to corroborate the efficacy of our proposed framework.