The wide spread of location-based social networks brings about a huge volume of user check-in data, whichfacilitates the recommendation of points of interest （POIs）. Recent advances on distributed representation she...The wide spread of location-based social networks brings about a huge volume of user check-in data, whichfacilitates the recommendation of points of interest （POIs）. Recent advances on distributed representation shed light onlearning low dimensional dense vectors to alleviate the data sparsity problem. Current studies on representation learningfor POI recommendation embed both users and POIs in a common latent space, and users＇ preference is inferred basedon the distance/similarity between a user and a POI. Such an approach is not in accordance with the semantics of usersand POIs as they are inherently different objects. In this paper, we present a novel translation-based, time and locationaware （TransTL） representation, which models the spatial and temporal information as a relationship connecting users andPOIs. Our model generalizes the recent advances in knowledge graph embedding. The basic idea is that the embedding ofa 〈time, location〉 pair corresponds to a translation from embeddings of users to POIs. Since the POI embedding shouldbe close to the user embedding plus the relationship vector, the recommendation can be performed by selecting the top-kPOIs similar to the translated POI, which are all of the same type of objects. We conduct extensive experiments on tworeal-world data.sets. The results demonstrate that our TransTL model achieves the state-of-the-art performance. It is alsomuch more robust to data sparsity than the baselines.展开更多
基金This work was supported by the National Natural Science Foundation of China under Grant Nos. 61572376 and 91646206, and the National Key Research and Development Program of China under Grant No. 2016YFB1000603.
文摘The wide spread of location-based social networks brings about a huge volume of user check-in data, whichfacilitates the recommendation of points of interest （POIs）. Recent advances on distributed representation shed light onlearning low dimensional dense vectors to alleviate the data sparsity problem. Current studies on representation learningfor POI recommendation embed both users and POIs in a common latent space, and users＇ preference is inferred basedon the distance/similarity between a user and a POI. Such an approach is not in accordance with the semantics of usersand POIs as they are inherently different objects. In this paper, we present a novel translation-based, time and locationaware （TransTL） representation, which models the spatial and temporal information as a relationship connecting users andPOIs. Our model generalizes the recent advances in knowledge graph embedding. The basic idea is that the embedding ofa 〈time, location〉 pair corresponds to a translation from embeddings of users to POIs. Since the POI embedding shouldbe close to the user embedding plus the relationship vector, the recommendation can be performed by selecting the top-kPOIs similar to the translated POI, which are all of the same type of objects. We conduct extensive experiments on tworeal-world data.sets. The results demonstrate that our TransTL model achieves the state-of-the-art performance. It is alsomuch more robust to data sparsity than the baselines.
