Meta-path Based Link Prediction In Aligned Heterogeneous Social Networks

With the development of Internet, smart terminals have been popular, and people will join in different kinds of social networks. This makes the methods imperfect that recommends entities to users within a single social network. Recently in research fields, there appears a new network model, called aligned heterogeneous social networks that connects different social networks, which makes it possible to utilize multi-network data to do recommendation. However, the related researches of aligned heterogeneous social networks are still at the starting stage, focusing only on a few dataset types and their corresponding user-user link and user-location link prediction problems. Other kinds of link prediction or recommendation problem have not attracted enough attention. Especially, features or factors that are being considered are too simple, and methods for utilization of multiple attributes and feature processing with selection are not well-designed. The prediction space is limited so that it is not appropriate for recommendation conditions, and the predicted results do not have time causality. Therefore, this thesis proposes user-entity link prediction problem in aligned heterogeneous social networks, namely user-entity link prediction problem, which includes these following research points:Firstly, an aligned heterogeneous social networks link prediction framework based on meta-path feature is proposed with transformation of the problem into a classification one in machine learning and an automatic method for feature definition based on meta-path, then the features by defining multiple methods to calculate network edge weight is extended into an initial feature set, and part of features is selected to generate final prediction model.Secondly, a novel feature selection algorithm called 2-P SFGFS is proposed. This algorithm combines the advantages of filter model and wrapper model together to select a feature subset from the whole feature set. Experiment results show that our feature selection method generate models that perform better in prediction and have low model complexity and time consumption than the existing ones.Thirdly, an aligned heterogeneous social network entity recommendation prototype system is implemented based on the above process and algorithms, which includes the functions of an aligned heterogeneous social network dataset crawling, feature definition and auto-generation, edge weight calculation, feature computation, feature selection and recommendation model generation as well as a recommendation result visualization module.Afterwards, the prototype system is used to crawl enough dataset of Foursquare and Twitter and perform verification experiments on Foursquare-user-user link prediction, Foursquare-user-location link prediction and Twitter-follower-friend link prediction. As can be seen in the selected feature types, which not only includes the meta-path types of factors of time, space and 2 or 3-degree-friend, but also those generated by different weight calculation methods of anchor link, popularity and tag. We evaluate the model by the measures of precision, recall, F1, proving our model has better prediction performance than current related works or some baseline algorithms.