Mitigating Hidden Confounding Effects for Causal Recommendation

TL;DR

This work addresses hidden confounding in causal recommendations by introducing Hidden Confounder Removal (HCR), a front-door based framework that decomposes the causal effect of item features on user feedback through mediators. By learning two identifiable components, $P(m|u,i)$ and $P(l|i,m)$, via multi-task training, HCR can estimate $P(l|u,do(i))$ without observing the hidden confounder $V$. The approach is instantiated with multimodal backbones and evaluated on three real-world datasets, showing consistent improvements over both standard and deconfounding baselines, particularly for less-active users. The results support the practicality of front-door adjustment in recommender systems and open avenues for dynamically adapting to temporal confounding and broader mediator choices.

Abstract

Recommender systems suffer from confounding biases when there exist confounders affecting both item features and user feedback (e.g., like or not). Existing causal recommendation methods typically assume confounders are fully observed and measured, forgoing the possible existence of hidden confounders in real applications. For instance, product quality is a confounder since affecting both item prices and user ratings, but is hidden for the third-party e-commerce platform due to the difficulty of large-scale quality inspection; ignoring it could result in the bias effect of over-recommending high-price items. This work analyzes and addresses the problem from a causal perspective. The key lies in modeling the causal effect of item features on a user's feedback. To mitigate hidden confounding effects, it is compulsory but challenging to estimate the causal effect without measuring the confounder. Towards this goal, we propose a Hidden Confounder Removal (HCR) framework that leverages front-door adjustment to decompose the causal effect into two partial effects, according to the mediators between item features and user feedback. The partial effects are independent from the hidden confounder and identifiable. During training, HCR performs multi-task learning to infer the partial effects from historical interactions. We instantiate HCR for two scenarios and conduct experiments on three real-world datasets. Empirical results show that the HCR framework provides more accurate recommendations, especially for less-active users. We will release the code once accepted.

Figures (8)

• Figure 1: Causal graphs for illustrating the hidden confounding effect. $V$: hidden confounders, $I$: item features affected by $V$, $U$: user features, $L$: user preferences, $M$: mediators.
• Figure 2: Model architecture for HCR. The training stage and inference stage use blue arrows and red arrows, respectively.
• Figure 3: A specific causal graph under the general causal graph (Figure \ref{['fig:causalG']}) where the variables $\{C,Z\}$ constitute the mediator $M$. $V$: hidden confounders, $Z$: integrated features of user-item pair, $C$: click feedback, $L$: like feedback.
• Figure 4: (a) Causal graph with hidden mediator $M^\prime$. (b) Causal graph with hidden mediator $M^\prime$ and edge $V \rightarrow M^\prime$. $M$ denotes the mediator that is related to the matching process of recommendation and $M^\prime$ denotes hidden mediator that is independent to the user features. (c) A special causal graph with the measured confounder $V^\prime$. $V$ denotes the hidden confounder and $V^\prime$ denotes the measured confounder.(d) A special causal graph with edge $V \rightarrow M$
• Figure 5: Performance of HCR and CT in active and less-active user groups. (a) the absolute performance; and (b) the relative improvements of HCR over CT. "AU" and "LAU" are short for the active user group and the less-active user group, respectively. In (a), bars with slash and without slash corresponds to CT and HCR, respectively. Better viewed in color.