Sheaf-inspired transformations improve recommendation ranking and coverage

Sheaf-Inspired Structured Scoring Extensions of LightGCN for Collaborative Filtering

Information RetrievalComputational Engineering, Finance, and ScienceComputational GeometryMachine Learning

Summary

Recommender systems try to suggest things you might like by looking at connections between users and items. This paper explores a new way to tweak the scoring method in a popular recommendation model called LightGCN, inspired by a mathematical concept called cellular sheaves. The authors tested several new types of scoring that change how user and item information interacts without changing the main recommendation process. These new methods sometimes improved the quality of recommendations and how many different items got recommended, depending on the dataset. However, the improvements are moderate, and the work doesn’t prove that the math idea itself is directly responsible for the benefits.

What this means in practice

  • For e-commerce platform engineers: Enhance product recommendation systems by adding structured scoring functions on top of graph-based models to improve diversity and ranking quality.
  • For media streaming service developers: Increase coverage of recommended movies or songs by integrating pair-dependent embedding transformations into existing graph recommendation pipelines.

Authors

Ksenia Belkovskaya, Timur Lidzhiev, Maria Stebliy, Alexander Tarakanov

Abstract

Graph-based collaborative filtering learns user and item representations by propagating embeddings over a user--item interaction graph. We investigate a modular, sheaf-inspired scoring extension placed on top of a three-layer LightGCN backbone. The motivation comes from cellular sheaves, where local vector spaces are related by linear restriction maps; here, this idea motivates structured, pair-dependent transformations of user and item embeddings rather than a change to graph propagation. We compare the identity inner-product scorer, diagonal, rank-one low-rank, and full matrix residual scorers, and a parameter-matched residual MLP baseline (\texttt{mlp\_full}) on Amazon-Book, Gowalla, MovieLens-1M, and Yelp2018. We report ranking quality and catalog coverage at cutoffs 10 and 20. The selected structured variants show modest, dataset-dependent changes in ranking quality and sometimes substantial changes in coverage. Our method is inspired by sheaf geometry but does not construct a cellular sheaf or implement sheaf-Laplacian diffusion; the experiments therefore do not establish a causal benefit from sheaf geometry itself.