An Unbiased Recommendation Algorithm Based on Causal Inference for Automatically Removing Multiple Biases
Di Wang, Chaoqun Ma, Chen Xu, An Chen · 2025
As an important technology and means of information filtering, recommendation is considered as an effective tool to solve information overload problem. Combining machine learning and other technologies, the recommendation system models user interests, by collecting user behavior data, to carry out personalized item recommendation. Since the user behavior data is obtained by observation rather than experiment, the distribution of the collected training data is not balanced and having non-randomness loss phenomenon, commonly causing multiple biases between the training data and the ideal data distribution. However, the existing unbiased recommendation algorithms either focus on one or two specific biases, or integrate multiple bias elimination methods to remove a certain bias, lacking the universal ability to consider multiple biases or even unknown biases in data. To solve the problem of automatic removal of various biases in recommendation systems, this paper proposes an unbiased recommendation algorithm based on causal inference aiming at the multiple known biases in rating data. Compared to existing algorithms, the proposed algorithm can accurately model user interests and improve the model recommendation accuracy and robustness. The research of this paper has important research significance and application value for maintaining the fairness and impartiality of internet information service algorithms and promoting the healthy and orderly development of algorithm related industries.