Multi-Cause Deconfounding for Recommender Systems with Latent Confounders

TL;DR

This work tackles latent confounders in recommender systems by introducing MCDCF, a multi-cause deconfounding framework that separately learns substitutes for user-side and item-side latent confounders from user behavior data. It leverages two independent VAEs to recover $x_u$ and $x_i$ from multi-cause embeddings, then conditions the prediction on $U,I,X_U,X_I$ to debias the implicit feedback signal $C$, with a Bayesian ranking objective and ELBO regularization. Theoretical justification via backdoor adjustment and empirical validation on three real-world datasets demonstrate that MCDCF improves recommendation accuracy while providing stronger debiasing than state-of-the-art baselines and ablations. The results imply that explicitly modeling and removing both user- and item-side latent confounders can significantly enhance the reliability and fairness of recommendations without extra private data. The approach offers a practical, privacy-friendly path toward robust causal recommendations in complex, real-world settings.

Abstract

In recommender systems, various latent confounding factors (e.g., user social environment and item public attractiveness) can affect user behavior, item exposure, and feedback in distinct ways. These factors may directly or indirectly impact user feedback and are often shared across items or users, making them multi-cause latent confounders. However, existing methods typically fail to account for latent confounders between users and their feedback, as well as those between items and user feedback simultaneously. To address the problem of multi-cause latent confounders, we propose a multi-cause deconfounding method for recommender systems with latent confounders (MCDCF). MCDCF leverages multi-cause causal effect estimation to learn substitutes for latent confounders associated with both users and items, using user behaviour data. Specifically, MCDCF treats the multiple items that users interact with and the multiple users that interact with items as treatment variables, enabling it to learn substitutes for the latent confounders that influence the estimation of causality between users and their feedback, as well as between items and user feedback. Additionally, we theoretically demonstrate the soundness of our MCDCF method. Extensive experiments on three real-world datasets demonstrate that our MCDCF method effectively recovers latent confounders related to users and items, reducing bias and thereby improving recommendation accuracy.

Figures (8)

• Figure 1: Three causal graphs illustrate the assumptions underlying existing research. The variables are defined as follows: $U$ (user), $I$ (displayed items), $C$ (user feedback), $Z$ (item popularity), $S$ (user interest), and $W$ (user conformity). (a) The causal graph represents the traditional recommendation model; (b) the causal graph represents the PDA model; (c) the causal graph represents the DICE model.
• Figure 2: Two causal graphs illustrate multi-cause debiasing methods. The variables are defined as follows: $U$ (users), $I$ (displayed items), $C$ (user feedback), $A_{ui}$ (user exposure, i.e., the user's implicit interaction data with the item), $X$ (the set of latent confounders), $X_I$ (latent confounders for items), and $X_U$ (latent confounders for users). (a) The first causal graph represents the DeConfounder method; (b) the second causal graph represents our proposed MCDCF method.
• Figure 3: A partial causal graph illustrating the latent confounders between users and items. $u^{\prime}$: substitute representation of user $u$, $i^{\prime}$: substitute representation of item $i$, $i_{im}$: items user $u$ has interacted with, $u_{um}$: users who have interacted with item $i$.
• Figure 4: The overall architecture of MCDCF, which involves two core steps. First, the MCDCF learns substitute latent confounders $x_u$ and $x_i$ representing users and items from the user behaviour dataset. Second, $x_u$ and $x_i$ are combined with their corresponding users $u$ and items $i$, resulting in fused features $f$ and $k$ that target the positive sample data, followed by model optimisation. This process enables MCDCF to generate debiased recommendations.
• Figure 5: The IOU curves of all method on the Douban-Movie dataset.

Theorems & Definitions (2)

• Definition 1: problem definition
• definition 1: $d$-separation pearl2009causality