Skip to main content
Back

Group RC-DMC Beats Solo Recommenders by Learning What Groups Actually Want

Group RC-DMC unifies low-rank matrix completion with attention-based group modeling to recommend items to entire groups of users, not just individuals

Paper, Full Access
  • New framework: Group RC-DMC combines low-rank matrix completion with a Set Transformer to recommend items to groups of users.
  • Core problem: Most recommenders target individuals; group recommendations require resolving conflicting preferences over sparse rating data.
  • Key mechanism: Nuclear-norm proximal steps (singular value thresholding) enforce low-rank structure in the latent space during training.
  • Group modeling: A Set Transformer aggregator uses attention to weight group members' preferences dynamically, not just average them.
  • Results: Achieves lower group RMSE than WBF and AF baselines on MovieLens and Goodbooks, across small, medium, and large groups.
  • Availability: No public code or model released; paper is on arXiv and provides sufficient detail for reimplementation.

Most recommendation systems are built around a single question: what does this user want? But in the real world, people watch movies together, travel in groups, and eat out as families. Deciding what to recommend to a group is a fundamentally harder problem, and a new paper proposes a framework called Group Rank-Constrained Deep Matrix Completion (Group RC-DMC) that tackles it head-on by fusing classical matrix algebra with modern attention mechanisms.

The group recommendation gap

The standard approach to building a recommender system is collaborative filtering: you represent each user and each item as a vector of latent features, and you predict ratings by measuring how well those vectors align. Collaborative filtering employs explicit or implicit feedback to model user-item connections, and most methods are based on matrix completion techniques that recover missing values in user-item interaction matrices. The challenge is that in practice, the low-rank assumption is a critical premise for matrix completion in recommender systems, which speculates that most information in interaction matrices is redundant.

When you scale this to groups, two problems compound. First, it is useful to quantify uncertainty in the ratings that all members of a group would assign to the same item, enabling more informed decisions when individual preferences may conflict. Second, real-world rating data is extremely sparse: most users have rated only a tiny fraction of available items, making it hard to learn reliable representations in the first place.

Most existing group recommender systems sidestep this by using simple aggregation strategies, like averaging individual user preferences or taking the most conservative member's rating. These approaches are brittle, especially when group members have very different tastes.

Two worlds, one framework

Group RC-DMC is built on top of RC-DMC, a prior framework that enforces an explicit low-rank constraint on the learned user-item matrix. The key idea is that real preference data lives in a low-dimensional subspace: even if you have millions of users and thousands of items, the actual variety of taste patterns is much smaller. A direct way to promote low-rank matrices is through nuclear norm regularization, which is the L1-norm of the singular values vector.