Research map / Time series / Causal and generative time series
Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling
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.
Introduces Seahorse, a unified encode-evolve-decode framework and executable benchmark with raw-coordinate likelihood reporting for neural spatiotemporal point processes, evaluated on three real datasets and the HawkesNest synthetic entanglement suite.
Problem
Neural spatiotemporal point process (STPP) models are hard to compare because implementations differ in preprocessing, coordinate normalization, splits, likelihood conventions, and evaluation protocols, and aggregate leaderboards hide which inductive biases fail under which data regimes.
Approach
A four-stage decomposition (history encoding, state evolution, decoding, next-event prediction) that places intensity-based, density-based, and generative STPPs under one executable contract, with fixed raw-coordinate splits, log-Jacobian-corrected NLL reporting, and controlled synthetic entanglement diagnostics via HawkesNest.
Key results
- On real data, Attn-CNF has the best test NLL on COVID (-2.305) and Citibike (-6.444), while DeepSTPP is best on Earthquakes (2.705).
- On the HawkesNest entanglement suite, NSMPP achieves the lowest NLL at all levels L0-L3 (e.g., -2.724 at L0), but its ground-truth intensity correlation is substantially lower than AutoSTPP and DeepSTPP.
- No model family dominates uniformly across real datasets and controlled synthetic regimes.
Surprising findings
- NSMPP leads likelihood at every entanglement level yet has weak intensity recovery and does not lead temporal CRPS.
- Classical Hawkes-TvCNF remains competitive with neural methods on Citibike.
Limitations
- Headline comparison covers only a three-dataset 2D subset of the thirteen-dataset catalog.
- Some expensive neural presets are computationally fragile and were run with canonical or short-sweep configurations.
- Several generative and score-based models give only approximate likelihoods, limiting direct comparability.
Experimental setup
Datasets: COVID Earthquakes Citibike HawkesNest
Benchmarks: COVID Earthquakes Citibike HawkesNest entanglement suite
Metrics: test NLL (raw coordinates) temporal CRPS intensity correlation log-intensity RMSE rollout W1
Baselines: DeepSTPP AutoSTPP NSMPP NJSDE Jump-CNF Attn-CNF SMASH DSTPP THP RMTPP
Builds on: AutoSTPP DeepSTPP Neural STPP SMASH DSTPP
Closest papers by meaning
- Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events (NeurIPS 2026 · similarity 0.89)
- EasyTPP: Towards Open Benchmarking Temporal Point Processes (ICLR 2024 · similarity 0.88)
- Beyond Point Prediction: Score Matching-based Pseudolikelihood Estimation of Neural Marked Spatio-Temporal Point Process (ICML 2024 · similarity 0.88)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · similarity 0.87)
- Deep Continuous-Time State-Space Models for Marked Event Sequences (NeurIPS 2025 · similarity 0.86)
- Learning Spatiotemporal Dynamical Systems from Point Process Observations (ICLR 2025 · similarity 0.86)
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.