Modeling and analyzing bias in recommender systems from multi-views: context, topic and evaluation [article]

Jing Yuan, Technische Universität Berlin, Sahin Albayrak
2021
With the explosive growth of information on the Internet, recommender systems have been broadly applied in user engaged systems to efficiently discover items of potential interest. In order to automatically generate such "guess what you like" results and serve matching recommendations, advanced machine learning and data mining techniques are applied in recommender systems. However, the model based methods that learn from historical data often result in unavoidable bias in recommendations from
more » ... fferent perspectives. The filtering bubble reflected in the biased recommendations leads to the fact that recommended targets fall into a narrow range. The existing research towards the bias problem in recommender systems focuses mainly on the bias adjustment within a specific modeling phase. The comprehensive understanding and generic bias countering approaches are still missing. In this thesis, we research on the bias problem in recommender systems from multi-views, including contextual bias, content-level understanding of bias, and the evaluation bias. These bias phenomena are observed in specific application scenarios. The modeling, analysis and evaluation of bias are conducted accordingly. First, in the recommendation scenario of IPTV systems, recommendations are more sensitive to the contextual influence due to the video genres and airing schedule. Thus we conduct the research on modeling and countering contextual bias in this scenario. Second, digital news portals form a special recommendation scenario. The user impression or clicking behavior is heavily affected by the content-level bias understanding. Targeting on the gap between the article level popularity bias and the content-level understanding of bias, we research on the topical bias representation for news articles and their potential predicting power. In addition, in the algorithm recommendation scenario, the single objective evaluation leads to the overlook on the other measurement targets. Therefore, multi-objective evaluation and candidate expansion are [...]
doi:10.14279/depositonce-11998 fatcat:dw6wm2ftsrbttmvkhrjpunuxxu