Research map / Large language models / Agents, code and math
ExPERT: Modeling Human Behavior Under External Stimuli Aware Personalized MTPP
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 ExPERT, a Transformer Hawkes Process extension that models external stimuli and personalization (user and text-metadata embeddings) for next-event time and type prediction, evaluated on five datasets.
Problem
Marked temporal point process models of human action sequences ignore external stimuli (e.g., coupons, incentives, deadlines) and personalized context, and prior stimulus-aware models are domain-specific and do not scale.
Approach
Build on THP with a stimulus-aware query projection in attention (separate query matrices for user-event vs. stimulus-to-event), a type-based causal mask linking stimuli to affected event types, and additive user-feature and Sentence-BERT event-metadata encodings; external events are used in attention but excluded from the likelihood/prediction losses.
Key results
- Averaged over datasets, ExPERT gives 9.35% gain in type-prediction accuracy and 7.38% RMSE reduction over best baseline.
- On TRANSACTIONS-6M, ExPERT reaches 58.66% type accuracy and RMSE 2.5112 vs THP's 52.91% and 3.3118.
- On synthetic complex-stimulus data, ExPERT achieves 22.65% lower RMSE than THP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add external-stimuli-aware modeling to THP | ExPERT · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | RMSE: 0.1233 -> 0.0797; 0.1006 -> 0.0956; 2.2130 -> 2.0406; 0.7070 -> 0.6990; 3.3118 -> 2.5112 | Helped |
| Add external-stimuli-aware modeling to THP | ExPERT · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 53.49 -> 61.30; 64.63 -> 67.87; 59.31 -> 64.77; 64.05 -> 64.63; 52.91 -> 58.66 | Helped |
| Use user personalization only, without external-stimuli modeling (ablation) | THP · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 53.49 -> 51.78; 64.63 -> 64.91; 59.31 -> 60.75; 64.05 -> 64.05; 52.91 -> 53.57 | Mixed |
| Model external stimuli with a distinct attention mechanism (ablation) | THP · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 53.49 -> 57.28; 64.63 -> 64.87; 59.31 -> 63.48; 64.05 -> 64.37; 52.91 -> 56.18 | Helped |
| Add user personalization to external-stimuli modeling (ablation) | ExPERTExt · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 57.28 -> 61.20; 64.87 -> 67.54; 63.48 -> 63.98; 64.37 -> 64.37; 56.18 -> 57.03 | Helped |
| Add a type-aware causal mask (ablation) | ExPERTExt+P(User) · MOOC-5M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 67.54 -> 67.79; 64.37 -> 64.52; 57.03 -> 58.32 | Helped |
| Add event metadata (ablation) | ExPERTExt+P(User)+CM · MOODLE-0.5M, MOOC-5M, SO-10M, DUNNHUMBY, TRANSACTIONS-6M | Type prediction accuracy: 61.20 -> 61.30; 67.79 -> 67.87; 63.98 -> 64.77; 64.52 -> 64.63; 58.32 -> 58.66 | Helped |
| Increase the number of external stimuli per sequence | ExPERT · Synthetic data, 1 to 10 external stimuli | Log-likelihood | Helped |
| Increase self-excitation relative to external-stimulus influence | ExPERT vs THP · Synthetic data, α from 0.3 to 0.8 | Log-likelihood | Mixed |
| Model combined starting effects, periodic stimulations, self-excitation, and deadlines | ExPERT vs THP · Synthetic complex-stimulus experiment | RMSE: 22.65% lower RMSE | Helped |
Surprising findings
- Authors report that external-stimulus modeling contributes more to gains than personalization.
- SSHP performs poorly on time prediction and is evaluated on sequence clusters due to scaling issues.
Limitations
- Modeling of external-stimulus generation dynamics is out of scope; external events are excluded from the likelihood.
- Gains are marginal on DUNNHUMBY, where external events are sparse.
- Requires domain knowledge to define the type-level causal mask.
Experimental setup
Datasets: MOODLE-0.5M MOOC-5M SO-10M DUNNHUMBY TRANSACTIONS-6M
Benchmarks: MOODLE-0.5M MOOC-5M SO-10M DUNNHUMBY TRANSACTIONS-6M
Metrics: RMSE type prediction accuracy log-likelihood
Baselines: SSHP RMTPP Intensity RNN SAHP NHP AttNHP THP
Builds on: Transformer Hawkes Process SSHP Sentence-BERT AttNHP
Closest papers by meaning
- Hyper Hawkes Processes: Interpretable Models of Marked Temporal Point Processes (NeurIPS 2026 · similarity 0.85)
- Deep Continuous-Time State-Space Models for Marked Event Sequences (NeurIPS 2025 · similarity 0.85)
- EasyTPP: Towards Open Benchmarking Temporal Point Processes (ICLR 2024 · similarity 0.85)
- Decoupled Marked Temporal Point Process using Neural Ordinary Differential Equations (ICLR 2024 · similarity 0.85)
- DanmakuTPPBench: A Multi-modal Benchmark for Temporal Point Process Modeling and Understanding (NeurIPS 2025 · similarity 0.84)
- TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation (CVPR 2025 · similarity 0.83)
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.