![[2602.17893] COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.17893] COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models
Summary
The paper presents COMBA, a novel approach for learning large graphs using state space models, emphasizing cross batch aggregation and graph context gating to enhance performance.
Why It Matters
As large graph data becomes increasingly prevalent in various applications, efficient learning methods are essential. COMBA addresses the challenges of applying state space models to graph structures, potentially improving computational efficiency and accuracy in machine learning tasks involving large graphs.
Key Takeaways
- COMBA introduces graph context gating to optimize neighbor aggregation in large graphs.
- Cross batch aggregation allows for scalable training of graph neural networks (GNNs).
- Theoretical analysis shows that COMBA reduces error compared to traditional GNN training methods.
- Experiments demonstrate significant performance improvements over baseline approaches.
- Public access to code and benchmark datasets will facilitate further research and application.
Computer Science > Machine Learning arXiv:2602.17893 (cs) [Submitted on 19 Feb 2026] Title:COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models Authors:Jiajun Shen, Yufei Jin, Yi He, xingquan Zhu View a PDF of the paper titled COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models, by Jiajun Shen and 3 other authors View PDF HTML (experimental) Abstract:State space models (SSMs) have recently emerged for modeling long-range dependency in sequence data, with much simplified computational costs than modern alternatives, such as transformers. Advancing SMMs to graph structured data, especially for large graphs, is a significant challenge because SSMs are sequence models and the shear graph volumes make it very expensive to convert graphs as sequences for effective learning. In this paper, we propose COMBA to tackle large graph learning using state space models, with two key innovations: graph context gating and cross batch aggregation. Graph context refers to different hops of neighborhood for each node, and graph context gating allows COMBA to use such context to learn best control of neighbor aggregation. For each graph context, COMBA samples nodes as batches, and train a graph neural network (GNN), with information being aggregated cross batches, allowing COMBA to scale to large graphs. Our theoretical study asserts that cross-batch aggregation guarantees lower error than training GNN witho...
