Research On User Social Relationship Prediction Based On Location

With the development of the GPS satellite navigation systems and the rising popularity of smart phones,applications on location-based social networks(LBSNs)have attracted tens of millions of users.In LBSNs,users can share their location information(i.e.,check-in)when they find a new place or take part in social activities.Huge volume of check-in data provides an opportunity for researchers to study users' various social behaviors.Among them,predicting user's friendship based on their check-in data has become a research hotspot.Therefore,researchers have carried out extensive research.This paper focuses on how to mine effective features from the users' check-in data to improve the performance of friendship prediction methods and proposes three location-based user social friendship prediction methods.The major contributions of this paper are as follows:· Friendship Inference in Location-Based Social Networks via Spatiotemporal Features(STIF)The location entropy has been widely used in the past studies,however,it can not dynamically compute the place popularity with the change of time period.In order to solve this problem,STIF proposes a fine-grained lo-cation weight feature.What's more,STIF also designs 12 features from three as-pect: fine-grained temporal features,weekday and weekend check-ins features and co-occurrence distance features for comprehensively describe the similari-ties between friend pairs and the differences between stranger pairs.· Combining implicit and explicit features for friendship inference in location-base social networks(CIFEF)Most of the past studies,including STIF,relies on a large number of feature engineering.To reduce the dependence on feature engineering,this paper proposes a method that combining implicit and explicit features for inferring friendship in LBSNs(CIFEF).CIFEF divides the trajecto-ry sequence of each active user into weekday trajectory and weekend trajecto-ry based on consideration that the same user has different trajectory patterns on weekdays and weekends.Then,it utilizes skip-gram model to learn the em-bedding vector of weekday location and weekend location,respectively.Besides,CIFEF also design an explicit feature named twcle to complement implicit fea-tures,the twcle introduces the time interval between two users' check-in into location entropy to measure the importance of user pair's common place.· Predicting friendship based on siamese network in location-based social net-works(simFriends)To overcome the CIFEF method can not capture the sequen-tial relation of the locations in user's check-in trajectory,this paper proposes a method that predict friendship based on siamese network in LBSNs(simFriend-s).Concretely,LSTM is used to directly model the user's check-in trajectory be-cause LSTM can well capture the sequence of check-in location.In addition, simFriends constructs a "user-place" bipartite graph from the user's check-in data.The user sequence that indirectly related to each user was obtained by using random walk in the bipartite graph,and then the user sequence is mod-eled by embedding technology.Finally,following the idea of siamese network,simFriends calculate the euclidean distance of two parts of vectors,respectively,and the contrast loss is used as the loss function to train the network model.Finally,a large number of experiments were performed on two public datasets and compared with state-of-the-art baseline methods,experiment result shows that the effectiveness and superiority of three methods proposed in this thesis.