A Clickbait Identification Method for Unbiased Recommendation

Clickbait often ignores facts and fabricates false information to attract clicks and traffic, undermining the credibility of the media industry. Unbiased has been proven effective in identifying exposure bias of clickbait. However, there are still two major challenges. First, when the model pushes news to users, the issue of similar news cannot be ignored. Second, existing methods mostly focus on introducing auxiliary information to effectively compensate for the lack of interactive information. To address these issues, in this paper, a Clickbait Identification Method for Unbiased Recommendation, namely CI-UR, is proposed. CI-UR (1) proposes a deduplication method that utilizes a relevance function between news titles and content to solve the similar news push. (2) can reduce the exposure bias of news and achieve unbiased recommendation without adding auxiliary information. Firstly, a dynamic sampling strategy is adopted to obtain user embedding vectors and news embedding vectors. Then, CI-UR uses a nonlinear model to obtain prediction scores, and constructs an unbiased method to remove the impact of false positive noise on model training during user interaction. Experimental results on two real-world news datasets show that CI-UR is effective in improving the accuracy of recommendation compared to state-of-the-art algorithms.