Neural Graph with Multifaceted Effects of Deep Latent Feature Representation for Point-of-Interest Recommendations

Community-contributed geotagged photos (CCGPs) have become an essential source for improving the quality of point-of-interest (POI) recommendations. CCGPs comprise photo albums associated with heterogeneous metadata that reflect user interests and POI properties. However, existing methods exploit only visual content and ignore context information. Moreover, they resort to matrix factorization, which combines latent user and POI features using an inner product. In this study, we proposed a neural graph to collaboratively learn a deep representation of multifaceted effects to achieve accurate and personalized POI recommendations. Visual content is used to learn the latent feature representation of implicit feedback and uncover two-fold homophily effects. Heterogeneous metadata are used to extract context information (i.e., weather), which is used to understand contextual constraints and determine two-fold contextual effects. Experimental results on a real-world dataset demonstrate the effectiveness of the proposed framework.