[Paper Review] LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation

Xiangnan He, Kuan Deng+4 more|arXiv (Cornell University)|Feb 6, 2020

Recommender Systems and TechniquesComputer Science44 references413 citations

TL;DR

LightGCN simplifies GCN for collaborative filtering by removing feature transformation and nonlinear activation, using only neighborhood aggregation and layer-wise embedding mixing, yielding substantial gains over NGCF.

ABSTRACT

Graph Convolution Network (GCN) has become new state-of-the-art for collaborative filtering. Nevertheless, the reasons of its effectiveness for recommendation are not well understood. Existing work that adapts GCN to recommendation lacks thorough ablation analyses on GCN, which is originally designed for graph classification tasks and equipped with many neural network operations. However, we empirically find that the two most common designs in GCNs -- feature transformation and nonlinear activation -- contribute little to the performance of collaborative filtering. Even worse, including them adds to the difficulty of training and degrades recommendation performance. In this work, we aim to simplify the design of GCN to make it more concise and appropriate for recommendation. We propose a new model named LightGCN, including only the most essential component in GCN -- neighborhood aggregation -- for collaborative filtering. Specifically, LightGCN learns user and item embeddings by linearly propagating them on the user-item interaction graph, and uses the weighted sum of the embeddings learned at all layers as the final embedding. Such simple, linear, and neat model is much easier to implement and train, exhibiting substantial improvements (about 16.0\% relative improvement on average) over Neural Graph Collaborative Filtering (NGCF) -- a state-of-the-art GCN-based recommender model -- under exactly the same experimental setting. Further analyses are provided towards the rationality of the simple LightGCN from both analytical and empirical perspectives.

Motivation & Objective

• Demonstrate that common GCN components (feature transformation and nonlinear activation) offer limited benefit for collaborative filtering.
• Propose a simplified Graph Convolution approach focused on neighbor aggregation for user/item embeddings.
• Show empirical gains of LightGCN over NGCF under identical experimental settings.
• Analyze why a linear, concise design works well for recommendation tasks.

Proposed method

• Use ID embeddings for users and items and propagate them on the user-item interaction graph using a simple, symmetric normalization-based aggregation.
• Combine embeddings from all propagation layers via a uniform or simple weighted sum to form final user/item representations.
• Predict interactions via the inner product of final user and item embeddings.
• Train only the initial layer embeddings with Bayesian Personalized Ranking (BPR) loss and Adam optimization.
• Provide matrix formulation showing E^(k+1) = D^(-1/2) A D^(-1/2) E^(k) and final E as a sum of layer embeddings.

Experimental results

Research questions

• RQ1 Do feature transformation and nonlinear activation improve performance in CF when using GCNs?
• RQ2 Can a lighter model preserving only neighborhood aggregation outperform heavier GCN-based CF models like NGCF?
• RQ3 How does multi-layer LightGCN compare to single-layer MF and NGCF across standard datasets?
• RQ4 What is the role of layer combination in mitigating oversmoothing for long-range propagation in CF?
• RQ5 Is LightGCN easier to train and analyze while delivering competitive or superior performance?

Key findings

• LightGCN consistently outperforms NGCF across Gowalla, Yelp2018, and Amazon-Book datasets in recall@20 and ndcg@20.
• Removing feature transformation and nonlinear activation benefits NGCF, with further gains when both are removed (NGCF-fn outperforms NGCF).
• Increasing layers from 1 to 3 yields better performance with diminishing returns; 3 layers often provide strong results.
• Layer combination (summing embeddings from all layers) captures self-connection effects and improves representational power without extra complexity.
• LightGCN achieves lower training loss and better generalization than NGCF, with substantial relative improvements (e.g., up to ~16-17% in recall/ndcg on multiple datasets).
• The model’s linear, simplified design offers interpretability and training ease while surpassing state-of-the-art baselines in their experimental settings.