Social Context And Temporal-Spatial Data Based Personalized Location Recommendation

The widespread popularity of mobile intelligent devices gives a result that the social services are being rapidly transferred to mobile terminal. In addition, the location based social network(LBSN) has emerged. Benefitting from the location based services,LBSN has successfully attracted billions of users. The massive social and movement information provides the data support for the mining of user behavior patterns. To better support location based services, there is the existence of practical significance to mining user social behavior and giving personalized location recommendation.LBSN contains rich social and check-in information. Social information reveals the intrinsic relationships between users, while check-in information reflects the real movement traces in the offline world. Intuitively, analysis and mining of social and check-in information can help recommender systems to recognize user behavior patterns better and provide more effective location recommendations. This paper primarily focuses on the research on social context based personalized location recommendation with the goal of improving recommendation performance. The main contributions and contents are as follows:(1) Human geographical movement exhibits significant temporal patterns on LBSN and is highly relevant to the temporal state. In order to inject the temporal patterns into recommendation model, this paper presents an effort to infer temporal state specific social circle from available check-in data combined with social network data. To this end, this paper divides social relationships into several sub-networks, each of which concerning a specific temporal state so that the proposed model can recognize the fine-grained social effects.(2) This paper make an effort to explore more comprehensive social factors, involving individual preference, preference similarity, social trust and closeness degree. Meanwhile, this paper proposes several reasonable methods to measure these social factors by exploiting the check-in information. Finally, these social factors are simultaneously injected into recommendation model to enhance the accuracy.(3) This paper proposes a novel STS recommendation model which incorporates social effects and temporal-spatial patterns. This model releases the limit of “matrix sparseness and “cold start” to some extent, and improves the recommendation performance.(4) Utilize low-rank matrix factorization technique to learn latent user and location feature matrices. Once the latent matrices are leaned, the predicted user preference in each temporal state can be captured by matrix multiply. Meanwhile, this paper provides several aggregation strategies to generate a uniform predicted matrix. Finally, a series of experiments on publicly available data show the proposed STS model can provide effective location recommendations and achieve significantly improved performance.