Improving recommendation diversity without retraining from scratch

Diverse recommendations strongly correlate with increased sales diversity, perceived ease of use, and general user satisfaction with recommendation systems. However, many recommendation models focus only on maximizing recommendation accuracy. This can lead to a lack of diversity due to the feedback loop of recommending already popular items. Retraining the entire model to increase diversity can be expensive, time-consuming, and impractical. To address this, we propose a refinement strategy that uses reinforcement learning objectives to penalize non-diverse behavior. This allows us to improve the diversity of any pre-trained model without retraining it from scratch and needing the original training settings and labels. We evaluate our approach using three deep learning recommendation models on the Yoochoose, RetailRocket, and Movielens datasets. Our refinement scheme improves recommendation diversity by up to 5% while maintaining competitive recommendation ranking performance in metrics such as HitRate and NDCG.