Calibrating User Preferences for Cross-Domain Recommendation via Target-Guided Representation Mapping

Abstract

Cross-domain recommendation (CDR) uses auxiliary interactions from a source domain to mitigate data sparsity in a target domain. However, implicit feedback in the source domain may include false positives that distort user representations and transfer misleading preferences to the target domain. We refer to this problem as noisy preference transfer. To address this problem, we propose Denoising Cross-Domain Recommendation (DCDR), a model-agnostic framework for representation-level preference calibration in overlapping-user CDR. Under the assumption that target-domain user representations provide a comparatively more reliable and task-aligned signal, DCDR first selects high-confidence overlapping users according to their source-domain training losses. It then learns an explicit target-to-source representation mapping and combines the mapped target-domain representation with the source-domain representation to obtain a calibrated source-domain user representation. The resulting preference scores are used to reweight source-domain interactions, reducing the influence of interactions inconsistent with the target-guided preference estimate without directly discarding potentially useful data. We instantiate DCDR on three representative CDR backbones and evaluate it on six cross-domain recommendation tasks constructed from two public datasets. Experimental results show that DCDR consistently outperforms existing denoising methods across different datasets, cross-domain tasks, and backbone models. These results demonstrate that target-domain user representations can provide an effective auxiliary calibration signal for mitigating the negative impact of noisy source-domain interactions and improving the reliability of cross-domain recommendation.