Research map / Time series / Causal and generative time series
EasyTPP: Towards Open Benchmarking Temporal Point Processes
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 EasyTPP, an open benchmark and software library for neural temporal point processes, with unified datasets, evaluation programs, and PyTorch/TensorFlow implementations of nine TPP models.
Problem
Neural temporal point process models lack a standardized benchmark, so results are hard to compare and reproduce across datasets, evaluation protocols, and frameworks.
Approach
A central repository that unifies TPP datasets, evaluation metrics (log-likelihood, next-event time/type prediction, long-horizon OTD), batched thinning sampling and NLL computation, and modular PyTorch and TensorFlow implementations of popular neural TPPs under a shared training protocol.
Key results
- IFTPP achieves the best log-likelihood among compared models because it computes likelihood in closed form, while Monte Carlo-integrated models lag behind.
- No single model wins on next-event prediction; NHP is best on Retweet and Taxi, and attention-based models (SAHP, THP, AttNHP) are generally competitive on Amazon, Taobao, and StackOverflow.
- Torch and TensorFlow implementations differ by within about ±1.5% on prediction metrics, and reimplementations match original papers within about ±5%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IFTPP's closed-form log-likelihood evaluation rather than Monte Carlo integration | TPP models · Synthetic, Retweet, and Taxi | log-likelihood | Helped |
| Use FullyNN's intensity architecture, whose PDF does not integrate to 1 | FullyNN · Synthetic, Retweet, and Taxi | log-likelihood | Hurt |
| Use attention-based TPPs instead of non-attention methods | TPP models · Amazon, Taobao, and StackOverflow | next-event prediction | Mixed |
| Use NHP instead of other evaluated models | TPP models · Retweet and Taxi | next-event prediction | Helped |
| Use neural TPPs instead of classical MHP | TPP models · Most evaluation tasks | performance | Helped |
| Use AttNHP or THP for long-horizon prediction | TPP models · Retweet | OTD | Helped |
| Use THP for long-horizon prediction | TPP models · Taxi | OTD | Helped |
| Implement models in TensorFlow instead of PyTorch | Equivalent TPP implementations · Predictive ability | relative difference: [-1.5%, 1.5%] | No clear effect |
Surprising findings
- The classical multivariate Hawkes process is generally worse than neural TPPs, but NHP is a comparably strong baseline to attention-based models.
- All the best models make poor predictions of future event times, suggesting performance saturation across architectures.
Limitations
- Autoregressive, locally normalized models suffer from cascading errors in long-horizon prediction.
- Only one synthetic and five real-world datasets are currently hosted, and the FullyNN model supports only single-type sequences.
Experimental setup
Datasets: Synthetic Hawkes-1D Amazon Retweet Taxi Taobao StackOverflow
Benchmarks: Synthetic Hawkes-1D Amazon Retweet Taxi Taobao StackOverflow
Metrics: log-likelihood time RMSE type error rate optimal transport distance
Baselines: MHP RMTPP NHP SAHP THP AttNHP FullyNN IFTPP ODETPP
Builds on: Hawkes process RMTPP Neural Hawkes Process Transformer Hawkes Process Intensity-free TPP
Closest papers by meaning
- Deep Continuous-Time State-Space Models for Marked Event Sequences (NeurIPS 2025 · similarity 0.90)
- Interacting Diffusion Processes for Event Sequence Forecasting (ICML 2024 · similarity 0.89)
- DanmakuTPPBench: A Multi-modal Benchmark for Temporal Point Process Modeling and Understanding (NeurIPS 2025 · similarity 0.89)
- Seahorse: A Unified Benchmarking Framework for Spatiotemporal Event Modeling (NeurIPS 2026 · similarity 0.88)
- Residual TPP: A Unified Lightweight Approach for Event Stream Data Analysis (ICML 2025 · similarity 0.87)
- Decoupled Marked Temporal Point Process using Neural Ordinary Differential Equations (ICLR 2024 · similarity 0.87)
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.