A Privacy-Preserving Framework with Multi-Modal Data for Cross-Domain Recommendation

TL;DR

This work tackles data sparsity and privacy leakage in cross-domain recommendation by introducing P2M2-CDR, a framework that fuses multi-modal data with a disentangled encoder to learn domain-common and domain-specific user representations. A privacy-preserving decoder employs local differential privacy and contrastive learning (domain-intra and domain-inter) to obfuscate and align embeddings during inter-domain transfer. Extensive experiments on four Amazon-based CD scenarios show that P2M2-CDR outperforms state-of-the-art baselines while maintaining a controllable privacy-utility trade-off, with ablations validating the contribution of each component. The approach advances privacy-aware cross-domain recommendations by exploiting rich modalities (texts, visuals, reviews) and robust contrastive objectives, with potential extensions to federated settings for even stronger privacy guarantees.

Abstract

Cross-domain recommendation (CDR) aims to enhance recommendation accuracy in a target domain with sparse data by leveraging rich information in a source domain, thereby addressing the data-sparsity problem. Some existing CDR methods highlight the advantages of extracting domain-common and domain-specific features to learn comprehensive user and item representations. However, these methods can't effectively disentangle these components as they often rely on simple user-item historical interaction information (such as ratings, clicks, and browsing), neglecting the rich multi-modal features. Additionally, they don't protect user-sensitive data from potential leakage during knowledge transfer between domains. To address these challenges, we propose a Privacy-Preserving Framework with Multi-Modal Data for Cross-Domain Recommendation, called P2M2-CDR. Specifically, we first design a multi-modal disentangled encoder that utilizes multi-modal information to disentangle more informative domain-common and domain-specific embeddings. Furthermore, we introduce a privacy-preserving decoder to mitigate user privacy leakage during knowledge transfer. Local differential privacy (LDP) is utilized to obfuscate the disentangled embeddings before inter-domain exchange, thereby enhancing privacy protection. To ensure both consistency and differentiation among these obfuscated disentangled embeddings, we incorporate contrastive learning-based domain-inter and domain-intra losses. Extensive Experiments conducted on four real-world datasets demonstrate that P2M2-CDR outperforms other state-of-the-art single-domain and cross-domain baselines.

Figures (8)

• Figure 1: An illustration of earlier CDR methods (a) and P2M2-CDR (b). In comparison to earlier CDR methods, P2M2-CDR considers (1) introducing multi-modal data (review texts, textual, and visual features) to disentangle more informative domain-common and domain-specific features and (2) utilizing local differential privacy technology to protect user privacy.
• Figure 2: The overall framework of P2M2-CDR. It contains two modules: (1) Multi-Modal Disentangled Encoder, which first incorporates user-item rating matrix and multi-modal information w.r.t. review texts and visual and textual features to learn initial user and item representations and then disentangles user representations into domain-common and domain-specific embeddings. It contains multi-modal feature learning and domain disentanglement components. (2) Privacy-Preserving Decoder, which introduces local differential privacy (LDP) to safeguard user privacy. This module includes decoupled feature obfuscation, contrastive learning, and information fusion components.
• Figure 3: Visualization of user obfuscated disentangled embeddings in the scenario: Phone&Sport. (a) Red points represent obfuscated domain-common embeddings and blue points indicate obfuscated domain-specific embeddings in the source domain (Phone); (b) Red points represent obfuscated domain-common embeddings in the source domain (Phone) and green points show the obfuscated domain-common embeddings in the target domain (Sport).
• Figure 4: Visualization of user obfuscated disentangled embeddings in the scenario: Sport&Cloth. (a) Purple points represent obfuscated domain-common embeddings and gray points indicate obfuscated domain-specific embeddings in the source domain (Sport); (b) Purple points represent obfuscated domain-common embeddings in the source domain (Sport) and orange points show the obfuscated domain-common embeddings in the target domain (Cloth).
• Figure 5: Effect of different weight parameter $\alpha$ in scenarios: (a) Phone&Sport (b) Sport&Cloth.