7. Conclusion
In this paper, we proposed a new design notion of topicfocused similarity-based trust evaluation and trust propagation to rate trustworthiness of tweets and users in Twitter. Compared to existing methods, our approach has three advantages: (1) enabling context-based trustworthiness estimation to focus on credibility in a specific topic domain; (2) utilizing credible news reports to infer trustworthiness of tweets exhibiting contextual similarity in textual, spatial and temporal features; and (3) combining semantic and contextual information with social networking information for trustworthiness propagation. Experiments on Twitter event detection demonstrated that our method can effectively extract trustworthy tweets while excluding rumors and noise. In addition, a comparative performance analysis demonstrated that our method outperforms existing supervised learning schemes using tweets manually labeled or tweets generated based on keyword matching as the training set. This paper assumes persistent attack behavior, i.e., a malicious user attacks without disguise whenever it has a chance. In the future, we plan to consider more sophisticated attack behaviors such as random, opportunistic, and insidious attack behaviors [31–34] to further test the robustness of our topic-focused similarity-based trust evaluation scheme.
