Users Relationship Strength Prediction Based On Spatio-temporal Context Co-occurrence

Recently,users relationship strength prediction based on spatio-temporal data has become a hot topic for many researchers.Previous researches had mainly focused on contextaware or spatio-temporal co-occurrence,and the time context is seldom considered.This paper proposes a novel transformation idea that artfully merge spatiotemporal context and spatiotemporal co-occurrence together from multiple views,and further improves the prediction accuracy.Our work is mainly from three aspects.First,we propose multi-view relationship strength prediction based on spatiotemporal context co-occurrence.Second,we propose relationship strength prediction method based on view fusion.Third,an application solution for predicting user relationship strength is given.We use the SNAP open source dataset Brightkite and Gowalla as our experiments,and we split the dataset into training set,validation set and testing set.The main contributions of our work are listed as follows:(1)The multi-view context co-occurrence method is presented in this paper,which artfully transforms the relationship between users in spatiotemporal data into the synonymous word relationship in Natural Language Processing domain,which artfully realize the fusion between context and co-occurrence.And this method also considers the time context information.Our method firstly generates spatiotemporal context sequences from multiple views,and then use tool in the NLP domain to extract user context co-occurrence feature based on multi-views.The feature represents users' check-in time co-occurrence,space co-occurrence,time context and space context.Finally,we use machine learning techniques to predict the relationship strength based on multiple views.Experiments show that the recall of the best Day-Location view in multi-views is 10% higher than for the Brightkite dataset under the same precision,8% higher for the Gowalla dataset than the EBM method.(2)In this paper,we propose feature fusion(FF)based on context co-occurrence feature according to the characteristics of feature level.The FF method is based on the complementarity of two view feature,merge the two sets of feature,and then use machine learning techniques to train and predict the strength of the relationship.At the same time,this paper also gives a multi-view Decision Fusion(DF)method.Experiments show that the FF method improves 3.6% in AUC on the Brightkite dataset and 4.3%in AUC on the Gowalla dataset compared to the best day-Location view proposed in this paper.The FF method increases the AUC by 1.4% on the Brightkite dataset,and1.6% on the Gowalla dataset compared to the DF method.Moreover,the FF method is 6.1% higher on the Brightkite dataset and 2.4% higher on the Gowalla dataset than the current best method SCI.(3)This paper proposes an application framework for predicting the strength of social network relationships.The framework includes the following modules: The data storage and management module contains three sub-modules(data structure,data storage and data visualization);The data modeling module mainly modularizes the two fusion methods(FF and DF);The model evaluation module performs a comprehensive evaluation of performance of the forecasting method,and gives the LIFT,ROC and PR curve.Finally,the output the strongest user relationship strength pairs in descending sort.