Research map / Graph learning / GNN architectures

How Does Message Passing Improve Collaborative Filtering?

NeurIPS 2024 · Graph learning › GNN architectures · approach: Graph rewiring and scalable message passing · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Investigates why message passing helps collaborative filtering, finds gains mainly come from forward-pass neighbor representations and benefit low-degree users more, and proposes TAG-CF, a test-time single-step neighbor aggregation applied to matrix factorization embeddings.

Problem

Graph-based collaborative filtering uses message passing (MP) costly during training; it is unclear why MP helps CF and whether the explanations from graph learning tasks transfer to CF.

Approach

Decompose LightGCN's similarity into neighbor-information and gradient terms via ablations, and show BPR/DirectAU implicitly optimize a Laplacian smoothness term; based on this, apply a single MP step only at inference time to pretrained MF embeddings (TAG-CF), with an optional degree-cutoff variant (TAG-CF+).

Key results

• TAG-CF improves MF (DirectAU-trained) overall NDCG@20 by 7.1% on average and low-degree NDCG@20 by 9.1%, with up to 39.2% gains on low-degree users for MovieLens-1M.
• TAG-CF adds on average about 0.05x10^3 seconds, under 0.5% of MF training time, and achieves average rank 1.2 on NDCG among graph-based baselines.
• Ablation on LightGCN: removing neighbor information drops NDCG@20 by 14.5-29.7%, whereas removing gradient updates drops it only 0.1-4.1%.

Measured effects reported in the paper

Surprising findings

• Forward-pass neighbor representations account for most of the gain from message passing in CF, while gradient updates contribute only about 2%.
• Message passing in CF helps low-degree users more than high-degree users, contrary to the usual GNN finding.

Limitations

• Normalization hyperparameters m and n must be tuned, and the default LightGCN normalization does not always work for bipartite graphs.
• TAG-CF is not competitive when MF is trained with BPR due to high-variance embedding norms.
• Test-time aggregation can hurt high-degree users, which motivates the degree-cutoff variant TAG-CF+.

Experimental setup

Datasets: Amazon-book Anime Gowalla Yelp-2018 MovieLens-1M Internal

Benchmarks: Amazon-book Anime Gowalla Yelp-2018 MovieLens-1M Internal

Metrics: NDCG@20 Recall@20 NDCG@10 Recall@10 running time

Baselines: MF ENMF UltraGCN LightGCN NGCF LightGCL SimGCL SGL

Builds on: LightGCN BPR DirectAU GCN

Closest papers by meaning

• LLM Collaborative Filtering: User-Item Graph as New Language (AAAI 2026 · similarity 0.87)
• gOCCF: Graph-Theoretic One-Class Collaborative Filtering Based on Uninteresting Items (AAAI 2018 · similarity 0.87)
• Hgformer: Hyperbolic Graph Transformer for Collaborative Filtering (ICML 2025 · similarity 0.85)
• FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.85)
• Semantic Enhanced Heterogeneous Hypergraph Network for Collaborative Filtering (AAAI 2025 · similarity 0.84)
• Language Representations Can be What Recommenders Need: Findings and Potentials (ICLR 2025 · similarity 0.84)