Causal Inference in Recommendation Systems: Disentangling Interests and Debiasing Short Video Recommendations

This presentation reviews recent research on applying causal inference techniques to recommendation systems. It first motivates the need for causal methods to address various biases (e.g., popularity bias, data sparsity) and to improve interpretability, fairness, and robustness.

The work is organized into three main parts:

Disentangling User Interest and Conformity : Introduces the DICE framework (Disentangling Interest and Conformity with Causal Embedding) that learns separate representations for user interest and conformity using causal embeddings, contrastive learning, and curriculum learning. Experiments on multiple backbones show consistent performance gains and more interpretable embeddings.

Disentangling Long‑Term and Short‑Term Interests : Proposes CLSR (Contrastive Learning framework of Long and Short‑term interests for Recommendation) which models long‑term stable interests and short‑term dynamic interests via separate encoders, contrastive proxy labels, and adaptive weighting. Extensive experiments on e‑commerce and short‑video datasets demonstrate superior performance over baselines.

Debiasing Short‑Video Recommendation : Addresses duration bias in micro‑video recommendation by introducing the WTG (Watch Time Gain) metric and its ranking‑aware variant DCWTG. A model‑agnostic adversarial method called DVR (Debiased Video Recommendation) removes duration‑related features, achieving unbiased recommendations and better balance between short and long videos.

Each part includes detailed methodological descriptions, structural causal models, loss functions, and visualizations of learned embeddings. The experimental results consistently show that incorporating causal reasoning and disentanglement leads to higher accuracy, better fairness, and more explainable recommendations.

References to recent conference papers (WWW 2021/2022, MM 2022, TOIS 2024) are provided.