Exploiting the Benefits of Popularity Bias Based on Causal Graphs in Optimizing Recommender Systems
Wenjing Sun · 2024
A common issue with recommend systems is the popularity bias, leading to long-tailed effects on item distribution. The resultant distribution discrepancy between hot and long-tail items would not only inherit the bias but also amplify the bias, which contradicts personalized recommendations. Existing methods tend to overlook the benign popularity bias caused by differences in item quality or directly injecting item popularity scores into prediction. In order to utilize item quality information revealing the item's inherent property, we use the casual graph to separate the beneficial popularity bias from the harmful popularity bias along the time dimension. We found that item quality, representing inherent properties, remains stable and static, whereas item popular bias, influenced by recent item clicks, is highly time-sensitive. Based on the therapy, we propose a novel framework to capture the critical cause-and-effect relationships named Time-Dimension Separating framework(TDS), where we model a click by casual graph generated from three components: the static item quality, the dynamic popularity effect, and the user-item matching score returned by any recommendation model. During testing, we employ counterfactual inference to mitigate the impact of item popularity. Significantly, our approach modifies the training process of recommendation models, making it applicable to a wide range of existing methods. Experiments on the Ciao, Amazon-Music, and Douban-Movie show that TDS outperforms the state-of-the-art methods. Especially the precision on Ciao, our TDS has improved by 46.1 percent over the baseline MF.