[论文解读] STAR-GCN: Stacked and Reconstructed Graph Convolutional Networks for Recommender Systems

Jiani Zhang, Xingjian Shi+2 more|arXiv (Cornell University)|May 27, 2019

Recommender Systems and Techniques参考文献 34被引用 54

一句话总结

STAR-GCN通过堆叠的编码器-解码器GCN块和基于屏蔽的重构来学习低维的用户/物品嵌入，以应对冷启动并在传递式和归纳式评级预测中实现强大的性能。

ABSTRACT

We propose a new STAcked and Reconstructed Graph Convolutional Networks (STAR-GCN) architecture to learn node representations for boosting the performance in recommender systems, especially in the cold start scenario. STAR-GCN employs a stack of GCN encoder-decoders combined with intermediate supervision to improve the final prediction performance. Unlike the graph convolutional matrix completion model with one-hot encoding node inputs, our STAR-GCN learns low-dimensional user and item latent factors as the input to restrain the model space complexity. Moreover, our STAR-GCN can produce node embeddings for new nodes by reconstructing masked input node embeddings, which essentially tackles the cold start problem. Furthermore, we discover a label leakage issue when training GCN-based models for link prediction tasks and propose a training strategy to avoid the issue. Empirical results on multiple rating prediction benchmarks demonstrate our model achieves state-of-the-art performance in four out of five real-world datasets and significant improvements in predicting ratings in the cold start scenario. The code implementation is available in https://github.com/jennyzhang0215/STAR-GCN.

研究动机与目标

• Motivate improved user/item representation learning for recommender systems, especially in cold-start scenarios.
• Develop a scalable GCN-based framework that learns low-dimensional embeddings end-to-end without relying on one-hot inputs.
• Introduce masking and reconstruction to enable embedding recovery for unseen nodes and to regularize learning.
• Identify and mitigate label leakage in GCN-based rating prediction through a sample-and-remove training strategy.
• Demonstrate effectiveness in both transductive (full observed train graphs) and inductive (cold-start) settings.

提出的方法

• Use a stack of graph encoder–decoder blocks where each encoder aggregates multi-type (rating-level) neighbor information.
• Replace one-hot node inputs with learnable low-dimensional embeddings for users and items.
• Mask a fraction of input node embeddings during training and reconstruct them to enable cold-start generalization.
• Introduce a reconstruction decoder to recover input embeddings from encoded representations, acting as a multi-task regularizer.
• Apply a sample-and-remove strategy to delete sampled training edges from the graph during each training step to avoid label leakage.
• Provide end-to-end training with a loss that combines supervised rating prediction and reconstruction losses across L blocks.

实验结果

研究问题

• RQ1 Can STAR-GCN achieve state-of-the-art performance on transductive rating prediction across real-world datasets?
• RQ2 How well does STAR-GCN handle inductive/cold-start rating prediction for unseen users/items?
• RQ3 Does masking/reconstruction improve embedding quality and generalization compared to standard GCN-based predictors?
• RQ4 What training strategies are needed to avoid label leakage in GCN-based rating prediction?
• RQ5 What is the effect of architecture choices (stacked vs recurrent blocks) and input features on performance?

主要发现

• STAR-GCN achieves state-of-the-art performance on four out of five transductive datasets.
• In inductive settings, STAR-GCN consistently and significantly outperforms baselines.
• Masking and reconstructing input embeddings improves performance and enables cold-start generalization.
• The sample-and-remove training strategy effectively mitigates label leakage and boosts test RMSE.
• Recurrent versus stacked variants offer competitive performance with fewer parameters; external features give mixed gains.