Reproduce “Learning from Negative User Feedback and Measuring Responsiveness for Sequential Recommenders” on OpenResearch

About this paper

Yueqi Wang, Yoni Halpern, Shuo Chang, Jingchen Feng, Elaine Ya Le, Longfei Li, Xujian Liang, Min-Cheng Huang and 4 more

Researchers from Google and OpenAI developed a "not-to-recommend" loss function to integrate negative user feedback directly into sequential recommender training, alongside a counterfactual simulation framework to measure responsiveness. Live experiments showed up to a 2.44% reduction in dislike rate and a 0.61% increase in daily active users, while simulations demonstrated a 60.8% reduction in content similarity for disliked items.

Key insights

• Directly integrating explicit and implicit negative feedback into the sequential retrieval model's training objective via the "not-to-recommend" loss significantly improves user experience and engagement metrics on large-scale industrial platforms.
• The proposed "not-to-recommend" loss function offers superior gradient stability and more targeted learning from negative signals compared to alternative approaches like using negative-valued cross-entropy weights or REINFORCE in large item spaces.
• The counterfactual simulation framework provides a robust and causal method for quantifying how effectively recommender systems learn from and adapt to negative user feedback, allowing for direct evaluation of responsiveness improvements.

Summary by alphaXiv.