Learning to Alleviate Familiarity Bias in Video Recommendation

TL;DR

This work tackles familiarity bias in video recommendations by introducing LAFB, a lightweight, model-agnostic post-ranking debiasing framework that learns personalized factors from discrete and continuous familiarity signals to adjust the User Rating Prediction Score $s_{u,v}$ into a debiased score $s^{debias}_{u,v}$. Debiasing factors are estimated via two paths: discrete bucket means for low-dimensional features and a neural net $f(b; heta)$ for continuous features, with $Adj_b$ produced either as an empirical mean or by the network. The final score integration uses $Score_{u,v} = R(s^{debias}_{u,v}, X_{u,v})$, enabling seamless deployment without retraining the base model. Offline and online evaluations show that LAFB reduces familiar content exposure while increasing novel content share (Novel WT Share) and emerging-creator exposure, all while maintaining overall engagement. The approach is demonstrated in production, including YouTube post-ranking, highlighting practical impact on content diversity and long-tail discovery.

Abstract

Modern video recommendation systems aim to optimize user engagement and platform objectives, yet often face structural exposure imbalances caused by behavioral biases. In this work, we focus on the post-ranking stage and present LAFB (Learning to Alleviate Familiarity Bias), a lightweight and model-agnostic framework designed to mitigate familiarity bias in recommendation outputs. LAFB models user-content familiarity using discrete and continuous interaction features, and estimates personalized debiasing factors to adjust user rating prediction scores, thereby reducing the dominance of familiar content in the final ranking. We conduct large-scale offline evaluations and online A/B testing in a real-world recommendation system, under a unified serving stack that also compares LAFB with deployable popularity-oriented remedies. Results show that LAFB increases novel watch-time share and improves exposure for emerging creators and overall content diversity, while maintaining stable overall watch time and short-term satisfaction. LAFB has already been launched in the post-ranking stage of YouTube's recommendation system, demonstrating its effectiveness in real-world applications.

Figures (4)

• Figure 1: Performance under multi-discrete familiarity modeling.
• Figure 2: Performance under multi-continuous familiarity modeling.
• Figure 3: URPS score distributions across fine-grained familiarity levels (low, medium, and high), shown before (a) and after (b) debiasing.
• Figure 4: Comparison of Model Predictions.