Causal Inference for Bias Mitigation in Kuaishou Recommendation Systems

Recommendation systems inevitably suffer from biases such as popularity bias and selection bias, which affect both user experience and ecosystem health. User interaction logs, used as training data, inherit these biases, leading to biased models.

The talk begins with an introduction to causal inference, defining causality versus correlation, and describing three levels of causal questions: discovering relationships, estimating intervention effects (uplift), and counterfactual reasoning.

Common causal inference methods are surveyed, including inverse propensity weighting (IPW), matching, and causal forest. The focus then shifts to recent research on causal embedding, exemplified by DICE (WWW'21) and a Huawei Recsys'21 paper, which separate user interest and conformity signals using specially constructed datasets.

Several works are reviewed that address bias from a causal perspective: a KDD'21 paper that removes popularity bias by subtracting direct effects, a study on headline bias, and a SIGIR'21 paper that uses a back‑door adjustment to eliminate popularity influence while preserving its positive contribution during inference.

In Kuaishou, three practical applications are described:

Popularity debias: using a back‑door operator to break the Z→I edge in the causal graph, adjusting the conditional probability P(C|U,I,Z) and introducing a controllable popularity factor γ.

Causal representation decoupling: extending DICE to learn separate interest and conformity embeddings, constructing positive/negative samples based on likes and reward scores, and training with a multi‑task loss.

Video completion‑rate debias: defining length‑aware thresholds for positive samples, applying IPW based on completion rates, and demonstrating improved playback and completion metrics across video lengths.

Experimental results show increased exposure for medium‑tail items, higher completion rates, and controlled popularity effects, confirming the effectiveness of the causal debiasing pipeline.

The conclusion highlights three challenges: calibrating the degree of debiasing, building a unified debiasing system for industrial deployment, and addressing the information‑filter bubble by shifting from purely accommodative to guidance‑oriented recommendation strategies.

Q&A sections address validation of interest vs. conformity embeddings, potential negative impacts of over‑debiasing, and strategies for long‑tail item bias mitigation.