COMPASS model learns per-user spectral allocation for robust recommendations

Graph neural networks have become a popular backbone for recommender systems, but their performance often degrades when implicit feedback is noisy and user interactions are sparse. A new study addresses this weakness with COMPASS, a Contrastive Multi-Decomposition Approach with Social-Guided Signals, described in a regular paper published in the International Journal of Data Science and Analytics.

The method targets a well-known limitation of low-rank approximation of the interaction matrix. Such approximations require committing to a single truncation rank that is then applied identically to every user. The leading spectral directions are supported by many co-occurrences and are well determined, while the trailing directions carry less energy and are estimated from sparser evidence. A single truncation must therefore either discard the trailing directions, removing detail that distinguishes users, or retain them, admitting estimation noise for users whose histories cannot support that level of detail.

COMPASS replaces this fixed global choice with a learned, user-conditioned allocation over spectral components of differing reliability. A single decomposition of the normalized interaction matrix is partitioned into disjoint spectral bands, and a continuous gate conditioned on observed user activity determines how much of each band enters the representation of a given user. The allocation is obtained through a Gumbel-Softmax parameterization and is learned jointly with the recommendation objective rather than specified in advance.

Social relations enter the framework as a soft contrastive regularizer that encourages socially consistent representations.

Experiments on the Yelp, LastFM, and Douban datasets show that COMPASS consistently achieves competitive improvements over strong baselines, with gains of up to 6.69% in NDCG@20. Additional noise injection experiments indicate that the framework maintains stable performance under corrupted interactions, suggesting favorable robustness to noisy feedback. The work remains a computational study evaluated on benchmark datasets, and the reported gains are measured on those experimental settings rather than in deployed systems.

Subject of Research: A learned spectral-band allocation method with social contrastive regularization for robust graph-based recommender systems.

Article Title: A contrastive multi-decomposition approach with social-guided signals for robust recommender systems

Article References: Aliabadi, M., Abdollahpouri, A., & Moradi, P. (2026). A contrastive multi-decomposition approach with social-guided signals for robust recommender systems. International Journal of Data Science and Analytics, 22(1), Article 350. https://doi.org/10.1007/s41060-026-01320-5

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01320-5

Keywords: recommender systems, graph neural networks, low-rank approximation, spectral decomposition, contrastive learning, social signals, noisy implicit feedback, Gumbel-Softmax, gradient normalization, NDCG@20, Yelp, LastFM