Research map / Architectures and efficiency / RNNs and normalization
Neuro-Symbolic 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.
Proposes NS-TPP, a neural-symbolic temporal point process that learns temporal logic rules end-to-end via differentiable rule embeddings and a sequential covering algorithm, reporting much faster rule learning than TELLER and CLUSTER.
Problem
Discovering compact, interpretable temporal logic rules that explain irregular event occurrences, which existing methods either prespecify, learn non-differentiably, or run too slowly.
Approach
Represent predicates as fixed embeddings and rules as learnable rule embeddings that act as soft-attention-like filters selecting predicates and temporal relations; the intensity is built from these logic-informed features and trained by negative log-likelihood, with rules added one at a time via sequential covering and then jointly refined.
Key results
- On average NS-TPP achieves a 112-fold speedup over TELLER and CLUSTER, with rule-learning accuracy rising from 49% to 93%.
- On the synthetic Group-3 (20,000 samples) event time prediction, NS-TPP reaches MAE 24.8616, the lowest among the nine baselines (best baseline CLUSTER 26.0351).
- On Car-Following, NS-TPP achieves MAE 3.1614 versus 3.7255 for the best baseline CLUSTER; on LowUrine it achieves 2.3262 versus 2.3675 for CLUSTER.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NS-TPP instead of CLUSTER | Synthetic Group-3 event time prediction · Synthetic Group-3 (20,000 samples) | event time prediction MAE: 26.0351 -> 24.8616 | Helped |
| Use NS-TPP instead of CLUSTER | Car-Following event time prediction · Car-Following | event time prediction MAE: 3.7255 -> 3.1614 | Helped |
| Use NS-TPP instead of CLUSTER | LowUrine event time prediction · LowUrine | event time prediction MAE: 2.3675 -> 2.3262 | Helped |
| Use NS-TPP instead of TELLER and CLUSTER | Synthetic temporal logic rule mining · Synthetic datasets | rule-learning accuracy: 49% to 93% | Helped |
| Use NS-TPP instead of TELLER and CLUSTER | Synthetic temporal logic rule mining · Synthetic datasets | training speedup: 112-fold speedup | Helped |
Limitations
- Accuracy metric is strict: a rule counts as correct only if it exactly matches ground truth.
- Real-data rules rely on medical literature for validation rather than clinician evaluation.
- Synthetic evaluation uses a constrained setup (at most one rule per sample, 30 predicates).
Experimental setup
Datasets: Synthetic temporal rule datasets (Groups 1-3) Car-Following MIMIC-IV LowUrine
Benchmarks: Synthetic Group 1-3 datasets Car-Following MIMIC-IV LowUrine
Metrics: rule recovery accuracy running time MAE of rule weights event time prediction MAE log-likelihood
Baselines: THP RMTPP ERPP GCH LG-NPP GM-NLF TELLER CLNN CLUSTER
Builds on: Transformer Hawkes Process TELLER CLUSTER
Closest papers by meaning
- Explaining Point Processes by Learning Interpretable Temporal Logic Rules (ICLR 2022 · similarity 0.89)
- Neuro-Symbolic Hierarchical Rule Induction (ICML 2022 · similarity 0.89)
- NeSTR: A Neuro-Symbolic Abductive Framework for Temporal Reasoning in Large Language Models (AAAI 2026 · similarity 0.86)
- End-to-End Learning of LTLf Formulae by Faithful LTLf Encoding (AAAI 2024 · similarity 0.86)
- Evolving Minds: Logic-Informed Inference from Temporal Action Patterns (ICML 2025 · similarity 0.85)
- Latent Logic Tree Extraction for Event Sequence Explanation from LLMs (ICML 2024 · similarity 0.85)
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.