SOCIAL NETWORKS RUMOR DETECTION APPROACH BASED ON POST-LEVEL FEATURE FUSION

The existing rumor detection methods largely neglect the correlation between post semantics, post publishers and post propagation status, which lead to low detection rates. To solve this problem, this paper proposes a rumor detection approach PF-HAN based on post-level feature fusion. The model used a Bi-LSTM with attention mechanism to generate the semantic representation of each post, and spliced it with the social network features of the corresponding post to preserve the correspondence between them. The integrated representation of the posts obtained by the fusion was input into the hierarchical attention network in the form of sequence to extract the temporal features and generate the final event representation for rumor discrimination. Experimental results over Sina Weibo and Twitter show that the F1 value of the model reaches 0.956 and 0.740 when the model performs the rumor detection task and it can complete the early rumor detection task with high accuracy.