Collaborative topological filtering with multi-hop recurrent pathological aggregation

Abstract Learning vectorial representations of users and items from their interaction data is a core approach for Collaborative Filtering. While a user’s or an item’s representation is commonly built upon low-hop features such as IDs and interaction history, some recent works argue the existence of higher-hop interactions, thereby motivating the use of multi-hop topological knowledge in representation learning. However, existing methods in this area explore only the local pathological connections and thus ignore the overall semantics along paths. To this end, this paper introduces a new CF approach that learns to explicitly inject the multi-hop topological features of a user or an item as a whole into its representation in an end-to-end manner. Specifically, we explore the multi-hop topology via the paths connecting a user or an item to its neighbors at different hops. To capture the entire topological information, we seamlessly integrate aggregator function with a recurrent neural network to jointly extract salient neighborhood information and detect the pathological semantics. We develop two neural network models, DF-CTF and DW-CTF, where the former focuses on modeling each individual path and the latter focuses on adapting to the path entanglement in multi-hop structures. Furthermore, we evaluate our proposed approach on three real-world benchmark datasets and demonstrate its superior performance against state-of-the-art methods.