Cornell-Abel ITPD runs 40.9x faster than PCMCI at 2,000 nodes
TL;DR
- ITPD ran 2.3x, 9.1x, and 40.9x faster than iterative IAMB, DF-PC, and PCMCI respectively on 2,000-node graphs while holding F1 at 0.997–0.998.
- On nonstationary graphs with 5% of edges randomly redrawn each timestep, ITPD scored F1 0.984–0.994 versus 0.562–0.808 for PCMCI, DYNOTEARS, and CD-NOD.
- The authors formalized soundness and completeness proofs in Lean 4.3, disclosed in Appendix D, alongside AI-assisted autoformalization using Claude, GPT-6, and Gemini.
A new Hugging Face paper from Cornell Tech, Abel AI, and Arizona State introduces Iterative Temporal Parent Discovery (ITPD), a constraint-based algorithm the authors describe as "a sound and complete constraint-based procedure for causal structure learning on multivariate time series with causal nonstationarity." On 2,000-node graphs, ITPD ran "2.3×, 9.1×, and 40.9× faster than iterative IAMB, DF-PC, and PCMCI respectively," while holding F1 between 0.997 and 0.998.
The nonstationarity test is the sharper number. With 5% of edges randomly redrawn at each timestep, ITPD held F1 at 0.984–0.994; PCMCI, DYNOTEARS, and CD-NOD sat at 0.562–0.808. Lead author Jacqueline Maasch (Cornell Tech) and co-authors Shu Wan, Hank Zhu, Zhiguang Wang (Abel AI) and Huan Liu (Arizona State) also published machine-verified soundness and completeness proofs in Lean 4.3, with AI-assisted autoformalization disclosed in Appendix A.
The paper is explicit about where it does not win. On DREAM4 gene regulatory networks at early timesteps with M=200, PCMCI, CD-NOD, and DYNOTEARS outperformed ITPD, and the authors list "lack of public real-world benchmarks" as a direct limitation on their evaluation. The method also assumes no contemporaneous links and no latent confounding after t=1. This is the latest paper flagged by our generative-AI tracker, which has logged 387 stories in the last 90 days.
Originally reported by huggingface.co
Read the original article →Original headline: Cornell-Abel ITPD Learns Nonstationary Time-Series Causal Graphs 40x Faster Than PCMCI at 2,000 Nodes