CTDG-SSM Beats Top Graph AI Models Using One-Tenth the Parameters
CTDG-SSM fuses HiPPO memory with graph Laplacian polynomials to achieve state-of-the-art on dynamic link prediction with 10x fewer parameters than competitors
Takeaways
- CTDG-SSM is a new state-space model for continuous-time dynamic graphs that simultaneously handles long-range temporal and spatial dependencies -- something no prior model achieved.
- The core innovation, CTT-HiPPO, extends the HiPPO memory framework to graphs by projecting compressed history through a polynomial of the graph Laplacian, making memory topology-aware.
- On the MOOC dataset (a key long-range temporal benchmark), it achieves 99.00% AUC-ROC vs. 89.58% for DyGMamba and 88.08% for DyGFormer.
- On a spatial long-range sequence classification task at chain length 20, CTDG-SSM scores 98.60% while DyGFormer scores 42.79% and DyGMamba 45.29%.
- The model uses roughly one-tenth the parameters of competing methods and trains 6x faster per epoch than DyGFormer on several datasets.
- Code is available; the paper is under review at ICLR 2026 (arXiv:2606.04672).
Most real-world networks are not static. Friendships form and fade, financial transactions flow, students engage with course content at irregular intervals. Modeling these systems as continuous-time dynamic graphs (CTDGs) -- where each interaction is a timestamped event rather than a snapshot -- gives far richer signal than the traditional approach of slicing time into discrete windows. But getting a neural network to actually learn from that richness has been stubbornly hard, especially when the patterns that matter span long time horizons and multiple hops across the graph. A new paper, "Learning Long Range Spatio-Temporal Representations over Continuous Time Dynamic Graphs with State Space Models", proposes a framework called CTDG-SSM that attacks both problems simultaneously -- and does it with a fraction of the parameters of competing methods.
The two walls every CTDG model hits
Existing approaches still face two primary challenges: (1) long-range temporal dependencies (LRT) -- the ability to preserve and use node states and interactions over extended time horizons; and (2) long-range spatial dependencies (LRS) -- the ability to capture multi-hop structural interactions beyond immediate neighborhoods in dynamic graphs. These are not just academic concerns. These capabilities have made CTDGs increasingly important in domains including finance, healthcare, e-commerce, and social network analysis. In financial fraud detection, for instance, money laundering typically spans long transaction chains rather than isolated local interactions -- exactly the kind of pattern that existing models miss.
The field has split into two camps, each solving one problem while ignoring the other. Event-driven models like TGN and JODIE update node states at each interaction and are computationally lean, but they mainly capture short-term patterns and struggle to preserve long-range temporal context. Sequence-based models like DyGFormer and DyGMamba use Transformers or Mamba to explicitly target long-range temporal dependencies, but these methods pre-process temporal data by restricting attention to one-hop temporal neighborhoods, thereby significantly limiting their ability to maintain LRS. No existing model handles both at once.
Building memory that knows the graph
The core insight of CTDG-SSM is to marry two powerful ideas: the HiPPO memory framework and graph Laplacian polynomials. HiPPO (High-order Polynomial Projection Operators) is a mathematical framework originally developed for sequence modeling. HiPPO defines a general mathematical and algorithmic framework for online compression and representation of sequential data via optimal projection onto polynomial bases, underpinning state-of-the-art models in recurrent neural networks, state-space systems, and Gaussian processes for efficient long-range sequence modeling. In plain terms: instead of trying to remember every past event, HiPPO compresses history into a compact set of polynomial coefficients that optimally summarize what happened.