Learning Temporal Representations of Bipartite Financial Graphs

Dynamic Bipartite graph is naturally suited for modeling temporally evolving interaction in several domains, including digital payment and social media. Though dynamic graphs are widely studied, their focus remains on homogeneous graphs. This paper proposes a novel framework for representation learning in temporally evolving bipartite graphs. It introduces a bipartite graph transformer layer, a temporal bipartite graph encoder based on an attention mechanism for learning node representations. It further extends the information maximization objective based on noise contrastive learning to temporal bipartite graphs. This combination of bipartite encoder layer and noise contrastive loss ensures each node-set in the temporal bipartite graph is represented uniquely and disentangled from other node-set. We use four public datasets with temporal bipartite characteristics in experimentation. The proposed model shows promising results on the transductive and inductive dynamic link prediction task and on the temporal recommendation task.