Combining mitigation treatments against biases in personalized rankings: use case on item popularity

Historical interactions leveraged by recommender systems are often non-uniformly distributed across items. Though they are of interest for consumers, certain items end up therefore being biasedly under-recommended. Existing treatments for mitigating these biases act at a single step of the pipeline (either pre-, in-, or post-processing), and it remains unanswered whether simultaneously introducing treatments throughout the pipeline leads to a better mitigation. In this paper, we analyze the impact of bias treatments along the steps of the pipeline under a use case on popularity bias. Experiments show that, with small losses in accuracy, the combination of treatments leads to better trade-offs than treatments applied separately. Our findings call for treatments rooting out bias at different steps simultaneously.