Applying Causal Inference to Debias Recommendation Systems at Kuaishou

The presentation begins with an overview of bias in recommendation systems, explaining how user interaction logs introduce selection and popularity biases that affect model training and user experience.

It then introduces causal inference as a framework to distinguish true user interest from confounding factors, describing three levels of causal reasoning: correlation exploration, intervention effect analysis, and counterfactual reasoning.

Recent research examples are discussed, including the DICE method (causal embedding with interest and conformity components), Huawei's bias‑unbias embedding approach, KDD21's total‑effect vs. direct‑effect decomposition, and SIGIR21's back‑door adjustment for popularity bias.

The speaker details Kuaishou's own applications: (1) popularity debiasing using back‑door operators and a modified conditional probability formulation; (2) causal representation decoupling that learns separate interest and conformity embeddings via multi‑task training; (3) video completion‑rate debiasing that defines length‑aware thresholds and applies IPW weighting.

Experimental results show increased exposure for medium‑tail items, improved completion rates across video lengths, and better separation of embeddings, confirming the effectiveness of the causal methods.

In the concluding section, the speaker reflects on the limited adoption of causal inference in recommendation, the difficulty of defining treatments, the need to balance debiasing strength, and the challenge of addressing filter bubbles, followed by a Q&A addressing embedding validation, potential negative impacts of debiasing, and strategies for long‑tail items.