TL;DR
The MSDCR model breaks the traditional cross-domain recommendation (CDR) mold by assuming that all involved domains are data-sparse, rather than requiring one "rich" source domain to help a "sparse" target. By using adversarial learning to simultaneously separate unique domain interests and adapt shared ones, it effectively transfers "complementary" preferences (e.g., using movie tastes to predict book interests) to significantly boost performance across multiple domains at once.
What Problem Does This Paper Address?
Cross-domain recommendation (CDR) is a popular strategy for solving the "data sparsity" problem—the fact that most users only interact with a tiny fraction of available items. Traditionally, CDR works like a mentor-student relationship: a "rich" source domain (with lots of data) helps a "sparse" target domain.
However, the authors identify two major flaws in this approach:
- The Rich Domain Myth: In the real world, almost every domain is sparse. Finding a "dense" auxiliary domain to act as a source is often impossible.
- Over-reliance on Invariants: Most models focus only on "domain-invariant" preferences (things a user likes regardless of the medium). They ignore "complementary" preferences—specific interests in one domain that could reasonably hint at interests in another (e.g., a user who watches martial arts movies might be interested in martial arts novels, even if they haven't read any yet).
Methods
The authors propose Multi-Sparse-Domain Collaborative Recommendation (MSDCR). The architecture is built on three main pillars:
- Multi-Domain Separation Network (MDSN): To ensure that domain-specific preferences are truly unique, the model uses adversarial training. A discriminator tries to guess which domain a preference embedding came from, while the encoder tries to make those embeddings distinct enough that they represent the "uniqueness" of that specific domain.
- Gated Aspect Preference Enhancement (GAPE): This is the "transfer" engine. It uses a gating mechanism to decide how much information to pull from other domains. It looks for "complementary" signals—preferences that are observed in Domain A but are beneficial for inferring potential interests in Domain B.
- Multi-Domain Adaptation Network (MDAN): While MDSN focuses on differences, MDAN focuses on similarities. It uses another adversarial process to extract "domain-invariant" preferences that are indistinguishable across all domains.
Finally, the model integrates these enhanced specific preferences and invariant preferences into a multi-task learning framework, allowing all domains to be optimized simultaneously.
Datasets
The model was evaluated using real-world data from the Douban website, focusing on three domains:
- Movie: ~155k items, 0.075% sparsity.
- Book: ~165k items, 0.022% sparsity.
- Music: ~166k items, 0.023% sparsity.
The researchers constructed three dual-domain scenarios (Movie-Book, Movie-Music, Book-Music) and one triple-domain scenario (Movie-Book-Music) using a set of 800 common users who overlapped across all three.
Key Insights and Ideas
- Collaboration over Mentorship: Instead of one domain "teaching" another, MSDCR treats domains as peers. Sparse domains can actually strengthen each other if the transfer mechanism is sophisticated enough.
- The Power of "Aspects": Rather than trying to map specific items across domains, the model maps aspects (latent features like genre or style). This makes the transfer much more flexible.
- Adversarial Tug-of-War: The model uses adversarial learning for two opposite goals at once: making some features as different as possible (domain-specific) and other features as similar as possible (domain-invariant).
- More is Better: One of the most striking findings was that the model performed better in the triple-domain scenario than in the dual-domain ones. This suggests that as you add more sparse domains, the "comprehensive understanding" of the user becomes exponentially better.
Key Results
MSDCR consistently outperformed state-of-the-art baselines (including NCF, CoNet, and BiTGCF) across all metrics:
- Superiority in Sparsity: In the Movie-Music scenario, MSDCR achieved an NDCG@10 of 0.2815 (Movie) and 0.3027 (Music), significantly higher than the best-performing dual-target CDR baselines.
- Robustness: As the researchers manually increased data sparsity (removing up to 80% of training data), the performance gap between MSDCR and other models actually widened, proving its resilience in data-poor environments.
- Ablation Success: Removing either the GAPE module or the adversarial separation (MDSN) led to a noticeable drop in performance, confirming that both "uniqueness" and "complementarity" are essential.
Limitations and Future Work
While the paper demonstrates strong results, a few areas remain for future exploration:
- User Overlap: The model relies on "common users" who exist in multiple domains to act as anchors for the transfer. In many real-world datasets, the number of overlapping users might be even smaller than the 800 used here.
- Computational Overhead: Running multiple adversarial networks and gating mechanisms across $S$ domains increases training complexity. Scaling this to dozens of domains simultaneously might require further optimization.
- Cold-Start Users: The paper focuses on improving recommendations for users who have some history in multiple domains; it does not deeply address users who are brand new to the entire ecosystem.
Summary updated