Embedding Taxonomical, Situational or Sequential Knowledge Graph Context for Recommendation Tasks

Learned latent vector representations are key to the success of many recommender systems in recent years. However, traditional approaches like matrix factorization produce that capture global distributions a static recommendation scenario only. Such user or item do not background knowledge and customized concrete situational context sequential history events leading up it. This is fundamentally limiting restriction for tasks applications, since state can depend on a) abstract information, b) current c) related observations. An illustrating example restaurant scenario, where user’s assessment situation depends taxonomical information regarding type cuisine, factors time day, weather location subjective individual experience this preceding situations. situation-specific internal captured when using collaborative filtering approach, knowledge, nature an individual’s cannot easily be represented matrix. In paper, we investigate how well state-of-the-art exploit those different dimensions relevant POI tasks. Naturally, represent such as temporal graph compare plain graph, taxonomy hypergraph embedding recurrent neural network architecture context-dimensions rich information. Our empirical evidence indicates most crucial prediction performance, while harder exploit. they still have their specific merits depending situation.