Research map / Time series / Long-term forecasting
GG-SSMs: Graph-Generating State Space Models
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.
Proposes GG-SSMs, which build a dynamic minimum spanning tree over feature cosine dissimilarities and propagate SSM state along it, evaluated on ImageNet, optical flow, event-based eye tracking and six time series datasets.
Problem
Traditional SSMs and fixed scanning strategies (Mamba, VMamba) process high-dimensional data along predetermined 1D paths, which fails to efficiently capture non-local dependencies.
Approach
Dynamically construct a graph from feature dissimilarities and compute a minimum spanning tree with Chazelle's near-linear algorithm, then perform SSM state propagation along tree paths instead of handcrafted scans.
Key results
- GG-SSM-B reaches 84.9% top-1 on ImageNet-1K, 1% above VMamba-B (83.9%) at 14.1 GFLOPs.
- GG-SSM achieves Fl-all of 2.77 on KITTI-15 after finetuning, versus 3.65 for the prior best VideoFlow-MOF.
- On LPW, GG-SSM reaches 89.33/98.89/99.50% detection rates at p3/p5/p10 with 0.22M parameters, above VMamba+Mamba.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use dynamic graph-based state propagation with an MST | SSM for ImageNet-1K classification · ImageNet-1K | top-1 accuracy: 83.9 -> 84.9 | Helped |
| use GG-SSM instead of VMamba+Mamba | event-based pupil tracking · LPW event-based | p3 detection rate: 89.00 -> 89.33 | Helped |
| use GG-SSM instead of Retina | event-based pupil tracking · INI-30 validation | centroid error: 3.24 (± 0.79) -> 3.11 (± 0.80) | Helped |
| use GG-SSM instead of 3ET | event-based pupil tracking · Synthetic LPW validation | centroid error: 5.33 (± 1.59) -> 5.11 (± 0.98) | Helped |
| use GG-SSM instead of VideoFlow-MOF (MF) | optical flow estimation · KITTI-15 finetuning | Fl-all: 3.65 -> 2.77 | Helped |
| replace Chazelle MST with Kruskal MST (ablation) | GG-SSM · ImageNet-1K Tiny | top-1 accuracy: 83.6 -> 83.5 | No clear effect |
| replace Chazelle MST with Prim MST (ablation) | GG-SSM · ImageNet-1K Tiny | top-1 accuracy: 83.6 -> 83.4 | No clear effect |
| replace Chazelle MST with Kruskal MST (ablation) | GG-SSM · ETTh2 horizon 192 | MSE: 0.353 -> 0.356 | No clear effect |
| replace Chazelle MST with Prim MST (ablation) | GG-SSM · ETTh2 horizon 192 | MSE: 0.353 -> 0.358 | No clear effect |
Surprising findings
- Ablation shows Kruskal, Prim, and Chazelle MST yield nearly identical accuracy (within 0.1%), so the MST algorithm choice mainly affects runtime.
Limitations
- Complexity claims of O(L) rely on the MST being computed in near-linear time and are not fully verified empirically.
- Ablation of MST algorithms covers only ImageNet Tiny and ETTh2, and the paper does not report GPU counts for the runs.
- Evaluation of time series and eye tracking is limited to the chosen datasets and baselines.
Experimental setup
Datasets: ImageNet-1K FlyingChairs FlyingThings3D Sintel KITTI-15 LPW INI-30 Exchange Weather Solar-Energy ETTh2 Traffic
Benchmarks: ImageNet-1K Sintel KITTI-15 LPW INI-30 Exchange Weather Solar-Energy ETTh2 Traffic ETTm2
Metrics: top-1 accuracy end-point error Fl-all detection rate within p pixels centroid error MSE MAE MACs parameters FLOPs
Baselines: VMamba Mamba S-Mamba iTransformer PatchTST Swin Transformer DeiT S4ND-Conv Vim RAFT FlowFormer++ MemFlow
Builds on: Mamba VMamba S4ND Chazelle MST
Closest papers by meaning
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.86)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.85)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.85)
- Spectral State Space Model for Rotation-Invariant Visual Representation Learning (CVPR 2025 · similarity 0.85)
- DAMamba: Vision State Space Model with Dynamic Adaptive Scan (NeurIPS 2025 · similarity 0.85)
- EyeGraph: Modularity-aware Spatio Temporal Graph Clustering for Continuous Event-based Eye Tracking (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.