Causal Deconfounding via Confounder Disentanglement for Dual-Target Cross-Domain Recommendation

TL;DR

CD2CDR addresses confounding in dual-target cross-domain recommendation by explicitly disentangling observed single-domain confounders (SDCs) and cross-domain confounders (CDCs) and applying backdoor adjustment to remove negative biases while preserving confounders' positive effects on predictions. It introduces a three-phase framework: Phase 1 disentangles user preferences into domain-shared, domain-specific, and domain-independent components to form comprehensive representations $E_u^*$; Phase 2 disentangles SDCs via a dual adversarial structure and CDCs via half-sibling regression to form confounder subspaces; Phase 3 applies backdoor adjustment with a confounder selection function and an MLP predictor to estimate $P(Y|do(Z))$ and predict interactions. The approach demonstrates strong empirical gains on seven real-world datasets, achieving average improvements of $6.17\%$ in $HR@10$ and $8.23\%$ in $NDCG@10$ over the best baselines (e.g., IPSCDR) in both dual-target CDR and cross-system scenarios. These results highlight the practical impact of explicitly handling observed confounders and integrating their positive effects into cross-domain knowledge transfer for more accurate and robust recommendations.

Abstract

In recent years, dual-target Cross-Domain Recommendation (CDR) has been proposed to capture comprehensive user preferences in order to ultimately enhance the recommendation accuracy in both data-richer and data-sparser domains simultaneously. However, in addition to users' true preferences, the user-item interactions might also be affected by confounders (e.g., free shipping, sales promotion). As a result, dual-target CDR has to meet two challenges: (1) how to effectively decouple observed confounders, including single-domain confounders and cross-domain confounders, and (2) how to preserve the positive effects of observed confounders on predicted interactions, while eliminating their negative effects on capturing comprehensive user preferences. To address the above two challenges, we propose a Causal Deconfounding framework via Confounder Disentanglement for dual-target Cross-Domain Recommendation, called CD2CDR. In CD2CDR, we first propose a confounder disentanglement module to effectively decouple observed single-domain and cross-domain confounders. We then propose a causal deconfounding module to preserve the positive effects of such observed confounders and eliminate their negative effects via backdoor adjustment, thereby enhancing the recommendation accuracy in each domain. Extensive experiments conducted on seven real-world datasets demonstrate that CD2CDR significantly outperforms the state-of-the-art methods.

Figures (6)

• Figure 1: Examples to depict single-domain confounder (SDC) and cross-domain confounder (CDC) du2022invariant.
• Figure 2: Causal graphs of dual-target CDR for deconfounding observed SDCs and CDCs. (a) Original causal graph. (b) Deconfounded causal graph after eliminating such observed confounders' negative effects by blocking backdoor paths via backdoor adjustment, as indicated by scissors wang2022causal.
• Figure 3: The structure of our CD2CDR Framework. The symbols and arrows not shown in the legend are defined in Table \ref{['tab:important_notation']} and Fig. \ref{['causal_graph']}.
• Figure 4: Average time comparison of confounder disentanglement phase across four tasks for CD2CDR and its variants.
• Figure 5: (a)-(b): Comparative performance analysis between CD2CDR and its seven variants with different backbones.