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

In this talk we review recent research on applying causal inference to recommender systems, covering three main topics: (1) causal embedding for disentangling user interest and conformity, (2) contrastive learning for separating long‑term and short‑term interests, and (3) debiasing short‑video recommendation by removing duration bias.

We first motivate the need for causal methods to address various biases, data sparsity, and to improve interpretability and fairness. A structural causal model is built to assign independent representations to interest and conformity, leading to the DICE framework that learns disentangled embeddings and uses multi‑task curriculum learning.

For sequential recommendation we propose CLSR, which models long‑term interest with a time‑invariant encoder and short‑term interest with a dynamic encoder, using proxy labels and contrastive loss to overcome the lack of explicit supervision.

In short‑video recommendation we identify duration bias and introduce the Watch‑Time‑Gain (WTG) metric and its ranking‑aware variant DCWTG. An adversarial debiasing method (DVR) removes duration information from the model output, achieving unbiased recommendations across various backbones.

Extensive experiments on public e‑commerce and short‑video datasets demonstrate that DICE, CLSR, and DVR consistently improve ranking metrics, provide more interpretable representations, and reduce bias compared with traditional baselines.

The work highlights how causal tools can enhance recommendation accuracy, explainability, fairness, and robustness in both static and sequential settings.