Causal Inference for Recommendation Systems: Disentangling User Interest, Conformity, Long‑Term/Short‑Term Interests, and Debiasing Short‑Video Recommendations

The talk introduces three research directions that leverage causal inference to improve recommender systems. First, a causal embedding (DICE) is proposed to disentangle user interest from conformity, assigning independent representations to each factor and using contrastive learning to handle bias and improve interpretability.

Second, the long‑term and short‑term interest disentanglement (CLSR) addresses the challenge of mixed interest signals in sequential recommendation. It builds separate encoders for stable long‑term interests and dynamic short‑term interests, uses proxy labels derived from historical pooling, and employs multi‑task curriculum learning to adaptively fuse the two representations.

Third, the short‑video bias mitigation work (DVR) defines a duration‑independent metric called Watch‑Time‑Gain (WTG) and its ranking‑aware variant DCWTG. An adversarial training scheme forces the recommender to ignore video‑duration features, achieving unbiased recommendations across videos of different lengths.

Extensive experiments on e‑commerce and short‑video datasets demonstrate that each method consistently improves ranking metrics, provides more interpretable representations, and reduces unfair exposure of long videos, confirming the effectiveness of causal‑based designs in recommender systems.