Causal Inference for Bias Mitigation in Kuaishou Recommendation Systems

Recommendation systems inevitably suffer from various biases, such as users preferring items at the top of the list or popular items, which leads to selection bias and feedback loops that degrade user experience and ecosystem health.

User interaction logs serve as training data, and directly using these biased logs introduces bias into the learned models, propagating biased recommendations.

To address this, causal inference is applied to identify and correct causal relationships between variables, distinguishing correlation from causation. Causal inference is described in three levels: (1) exploring correlations, (2) estimating the effect of interventions (e.g., uplift models), and (3) counterfactual reasoning to determine what would happen under alternative actions.

Typical causal methods include inverse propensity weighting (IPW), matching, and causal forest. Recent research highlighted includes:

DICE (WWW'21) – a causal embedding approach that separates user interest from conformity using triples.

Huawei’s Recsys'21 paper – splits feedback into biased and unbiased components and uses an information‑bottleneck loss.

KDD'21 work – removes popularity bias by treating it as a direct effect and subtracting the direct effect from the total effect.

SIGIR'21 work – applies a back‑door adjustment to eliminate popularity influence on item exposure.

In Kuaishou, three concrete applications were deployed:

Popularity debias – a back‑door operator removes the edge from popularity (Z) to item relevance (I) during training, while retaining the beneficial edge from popularity to content quality (C) during inference. The model predicts P(C|U,I,Z) as a product of a matching score and a popularity factor, with a controllable γ parameter and pair loss.

Causal representation decoupling – user behavior is split into interest and conformity embeddings. Positive/negative samples are constructed using reward (watch time + interactions) for interest and like count for conformity. Multi‑task training jointly optimizes feedback loss (main task) and the two embedding losses.

Video completion‑rate debias – a length‑aware threshold derived from the bimodal distribution of completion rates defines positive samples. IPW weights based on video length are applied to the loss to balance short and long videos.

Experiments show increased exposure and play counts for medium‑tail items, reduced exposure of low‑quality click‑bait, and higher completion rates across video lengths. However, challenges remain: determining the appropriate debiasing strength, building a unified debiasing system, and addressing information‑filter bubbles by shifting from purely satisficing recommendations to guiding user exploration.

The session concluded with a Q&A covering validation of interest vs. conformity embeddings, potential negative impacts of over‑debiasing, and strategies for long‑tail item correction.