AutoDebias based on NeuralCF

Abstract

Recommendation systems rely on user behavior data (ratings, clicks, etc.) to build personalized models. Nevertheless, the data collected was always observational rather than experimental, which can lead to a variety of biases. Data deviations will directly affect the operation results of the recommendation system. Users may not favor the items recommended by the system to them, so this is an imperative problem that must be resolved as soon as possible. The majority of current efforts to debias recommendation systems focus only on one or two specific biases and are not universally applicable. As a result, Chen et al. and their group developed a general debiasing framework (AutoDebias) containing most of the debias strategies, which overcomes the problem of low flexibility of most of today's debiasing techniques. However, the model still has room for improvement. Chen et al. chose Matrix Factorization (MF) as the benchmark model for AutoDebias. In our study, however, we found that NeuralCF model may be a better alternative to MF. Therefore, the goal of our work is to replace the MF base model in AutoDebias with the NeuralCF model, and to conduct relevant experiments in order to validate our hypothesis. According to experiments, when using NLL as evaluation metrics, NeuralCF has better performance than MF, which improved 14.3% in final results.