Theme-Aware Social Strength Inference from Spatiotemporal Data

Abstract

The popularity of location based services has resulted in rich spatiotemporal data that indicates whether persons have social connections. This valuable indication can be used in a wide range of applications such as friend recommendation in social networks or target advertisement for Internet companies. The state-of-the-art approach only considers people who visit non-popular locations together are more socially related. Further, none of existing methods notices that themes of co-occurrence behavior, e.g. dating, of every pair of persons can be used to infer their social strength. In this paper, we novelly introduce the theme to measure the social strength of two persons. A theme, mathematically, is in the form of a probabilistic distribution over Spatiotemporal Windows(SWs), the unit for co-occurrence. In this paper, we propose a Theme-Aware social strength Inference(TAI) approach that mines themes from co-occurrence behaviors consisting of SWs and trains each theme with its contribution to social strength. We employ tf-idf concept for SW and design a novel dynamic programming algorithm to find proper SWs. Extensive experiments are conducted on real dataset and the results show that our method can significantly improve the effectiveness, i.e. more than 5% to 15% in precision under the same recall over the state-of-the-art approach.