DPCRec-Dual Preferences Learning with Capsule Network for Next POI Recommendation

The goal of the next Point-of-Interest (POI) recommendation task is to predict a list of POIs that a user might visit next, based on their check-in history, including the locations and times of their check-ins. This list includes the probability of visiting each POI. Clearly, the performance of this task is closely related to how well the user's interests could be extracted. Previous research has made significant progress in this area, but there are still some challenges: 1) Current methods cannot fully extract user interests from different perspectives. 2) They struggle to effectively combine multiple or diverse user interests for prediction. In this paper, a Dual Interest Capsule Recommendation (DPCRec) is proposed to address these two issues. For the first problem, user interests are divided into long-term and short-term interests and introduce Advanced Long-Term Interest Feature Extractor(ALE) and Advanced Short-Term Interest Feature Extractor(ASE) to extract these interests separately. Among them, ALE uses a multi-head attention mechanism with deep attention and residual connections to capture long-term interests, while ASE employs a GRU module enhanced with attention and multi-gating techniques to capture short-term interests. For the second problem, a novel capsule network called Capsule Deep Pointer (CapsDP) is proposed by us, which effectively combines long-term and short-term interests and maps the hierarchical relationship between user interests and the POIs to be predicted. Extensive experiments on three datasets show that our model outperforms ten baseline models, achieving state-of-the-art results.