Aiming at the Target: Filter Collaborative Information for Cross-Domain Recommendation

TL;DR

This paper tackles negative transfer in cross-domain recommendation by explicitly preserving target-domain user similarity while transferring knowledge from a source domain. It introduces CUT, a backbone-agnostic framework with a TARGET phase that learns a target-driven user similarity matrix and a TRANSFER phase that uses a one-layer user transformation plus a contrastive regularization term to filter irrelevant source information. The method maintains target-domain user relations through $\mathbf{F}$ and $\mathbf{L}^c$, enabling more precise cross-domain knowledge transfer. Empirical results on six cross-domain tasks from two real-world datasets show substantial improvements over both single-domain baselines and prior cross-domain methods, with robustness to data sparsity and clear mitigation of negative transfer.

Abstract

Cross-domain recommender (CDR) systems aim to enhance the performance of the target domain by utilizing data from other related domains. However, irrelevant information from the source domain may instead degrade target domain performance, which is known as the negative transfer problem. There have been some attempts to address this problem, mostly by designing adaptive representations for overlapped users. Whereas, representation adaptions solely rely on the expressive capacity of the CDR model, lacking explicit constraint to filter the irrelevant source-domain collaborative information for the target domain. In this paper, we propose a novel Collaborative information regularized User Transformation (CUT) framework to tackle the negative transfer problem by directly filtering users' collaborative information. In CUT, user similarity in the target domain is adopted as a constraint for user transformation learning to filter the user collaborative information from the source domain. CUT first learns user similarity relationships from the target domain. Then, source-target information transfer is guided by the user similarity, where we design a user transformation layer to learn target-domain user representations and a contrastive loss to supervise the user collaborative information transferred. The results show significant performance improvement of CUT compared with SOTA single and cross-domain methods. Further analysis of the target-domain results illustrates that CUT can effectively alleviate the negative transfer problem.

Figures (5)

• Figure 1: User similarity distortion in cross-domain recommendation. For a target user $u$, some users have similar source-domain preferences and different target-domain preferences, e.g. $u_2$, which will lead to negative transfer of $u_2$ for $u$. Our CUT framework alleviates it by filtering misleading collaborative information from users with different target-domain preferences (i.e., $u_2$ and $u_4$).
• Figure 2: An overview of Collaborative information regularized User Transformation (CUT) framework. It includes a TARGET phase to learn user similarity and a TRANSFER phase to filter useful source information to transfer to the target domain. Key components of the TRANSFER phase include a user transformation layer and a specially designed contrastive loss.
• Figure 3: Performance on sparse target domain dataset. We randomly sample the target Amazon Sports training dataset interactions with different retain fractions, while fixing the source Amazon Cloth dataset and the target test dataset.
• Figure 4: Ablation study of the target-driven user similarity module and user representation transformation module in TRANSFER phase.
• Figure 5: Performance comparisons of the user history-based similarity and the backbone-driven similarity.