User Preference Learning based on Automatic Environment Classification for General Debiasing in Recommendation Systems

As a core technique to alleviate the problem of information overload, recommendation systems are commonly subject to various biases that lead to unsatisfactory recommendation results. Recently, several scholars have conducted studies on the bias problem and proposed some approaches to mitigate its negative effects. However, these methods are only effective for one particular bias and cannot be adapted to the actual situation where multiple biases interact with each other. To this end, we propose a general debiasing framework based on causal learning model with representing the confounding biases as a complicated environment. This framework assumes that a user's representation consists of his/her genuine preferences and confounding diverse preferences (i.e., the preferences influenced by different biases), it identifies the user's genuine preferences and diverse preferences based on the automatic classification of multifarious environments, so as to improve the recommendation performance via multi-view debiasing. Specifically, to obtain an unbiased user representation, we combine the invariant learning and the unsupervised learning approaches to capture the environmental factors that cause different biases. Finally, we test the proposed approach on two real-world datasets. Experimental results show that our method can not only achieve multi-view debiasing, but also improve the performance of the recommendation systems.