Integrating Sign Prediction With Behavior Prediction for Signed Heterogeneous Information Networks

People often use online social networks not only to express attitudes towards others, but also to make decisions, which forms signed heterogeneous information networks. Both sign prediction and behavior prediction can provide useful information for networks analysis, each of which has been a hot topic. However, existing methods for sign prediction mainly rely on the features from labeled links but ignore users' behavior and the features from unlabeled links, which often leads to dumb results. Similarly, inferring users' behavior without considering links' signs is dull as well. In order to solve this issue, in this paper, we present a novel model called SPBP to integrate Sign Prediction with Behavior Prediction in the context of signed heterogeneous information networks. It simultaneously captures users' social links (including both labeled links and unlabeled links) and users' behavior to improve the accuracy of prediction. First, due to the lack of labeled links in main stream social networks, we propose correlation estimation methods to estimate social correlation and behavioral correlation between users respectively. Then we encode structural balance-based features and status-based features according to social psychology theories. With the extracted features, we propose a sign prediction algorithm based on transfer learning to use knowledge extracted from related source networks to train the target network, which can effectively make up for the incompleteness of target samples. Finally, we propose a behavior prediction algorithm based on the predicted signs of links. Extensive experiments conducted on real-world signed heterogeneous information networks, Epinions, Slashdot and Wiki-RfA, demonstrate that SPBP can effectively solve both the sign prediction problem and the behavior problem.