Attention-based neural networks for trust evaluation in online social networks

Trust relationship prediction has gained a lot of interest in online social networks as it allows users to escape complex analysis in decision-making. However, most existing work fail to consider two kinds of complementary crucial user data, users' trusted neighbors and temporal continuity of user behaviors, simultaneously. In this paper, we analyze the complementarity between the crucial user data mentioned above and design an attention-based neural networks model, GainTrust, to aggregate the two kinds of user data for trust relationship prediction task. Specifically, we firstly map the users' trust neighbors and behavioral records into the same feature space by designing heterogeneous network embedding layer. Then, we develop multi-layer Long Short-Term Memory (LSTM) network to further learn the time series features of users over multiple time slots. Finally, after integrating time series features with trusted neighborhood features, we design two-level multi-head attention mechanism to obtain global trust features for trust evaluation among users. A series of extensive experiments are conducted on two real-world datasets and the results demonstrate the effectiveness of the proposed approach.