Social Tie Prediction Using Spatiotemporal Data

Predicting social tie between users using spatiotemporal data, which is used to predict whether there is social tie between any pair of users, is one of the important research directions, attracting numerous scholars’ research and exploration. We propose a new method to predict social tie, which can fully extract co-occurrence feature of users. The three main contributions are as follows:1) Traditional method mainly focus on user’ co-location feature but ignore user’ co-time feature. We propose a feature extracting method based on LDA, which can achieve not only co-location feature but also co-time feature simultaneously. Researches show that there would be larger probability of existing social tie between users who often check-in at the same place or at the same time. According to the frequency of each pair of user appear at each place, users’ topic distribution feature(i.e. co-location feature), are deeply extracted by LDA. The topic distribution feature is derived from co-occurrence among users and locations essentially. So, it fully describes the information of co-occurrence among users and locations. Similarly, according to the frequency of each pair of user appear at each period, users’ topic distribution feature,(i.e. co-time feature), are deeply extracted by LDA.(Precision, Recall) which social tie prediction based on LDA on Brightkite data set and Gowalla data set can reach is(72.6%, 72.7%) and(75.8%, 66.4%) respectively.2) Since time feature and space feature of user’ check-ins are not considered by LDA simultaneously, so this method is a coarse-grained feature extracting method. We propose a fine-grained method based on word vector algorithm, word2 vec, which considers both time feature and space feature of user’ check-ins simultaneously. By this method, co-locationtime feature and co-time-location feature are achieved. These kinds of features consider time feature and space feature simultaneously. Compared with LDA-based social tie prediction, on Brightkite data set, Precision and Recall rate increase 5.3% and 6.4% respectively, while on Gowalla data set, increase 11.9% and 10.4% respectively.3) In order to take full advantage of all co-occurrence feature to predict social tie, a decision fusion model is proposed. This model merges two kinds of social tie strength, achieving merged decision making feature: one is predicted by co-location feature, co-time feature which created by LDA model, and the other is predicted by co-location-time feature, co-time-location feature which created by word2 vec. Experiments show that on Brightkite data set, compared with LDA, Precision and Recall rate increase 7.1% and 8.2% respectively, compared with word2 vec, increase 1.8% and 1.8% respectively. On Gowalla data set, compared with LDA, Precision and Recall rate increase 14.8% and 13.0% respectively, and compared with word2 vec, increase 2.9% and 2.6%. These results in EBM show that EBM outperforms all existing models. On Gowalla data set,(Precision, Recall) of EBM is up to(80%, 70%).(Precision, Recall) of decision fusion model which is proposed in this paper is up to(90.6%, 79.4%). Therefore, compared with EBM, Precision and Recall rate increase 10.6% and 9.4% with our method.