Research map / Time series / Causal and generative time series
Neural Temporal Point Processes for Forecasting Directional Relations in Evolving Hypergraphs
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 DHyperNodeTPP, a sequential generative neural temporal point process that forecasts directed hyperedges in evolving hypergraphs by first predicting event nodes, then candidate hyperedges, then scoring them.
Problem
Forecasting higher-order directional relations (directed hyperedges) in temporal networks, where the number of possible hyperedges is exponential in the number of nodes, making exhaustive search infeasible.
Approach
A three-stage sequential generative process: a lognormal TPP node-event forecaster identifies nodes observing events, a Bernoulli candidate generator predicts projected adjacency and size vectors for those nodes, and a directed hyperedge predictor (cross- and self-attention scoring of left/right groups) selects true hyperedges from candidates; memory-based temporal node representations with batch
Key results
- On event type prediction (MRR), DHyperNodeTPP reaches 98.72 on Bitcoin and 79.01 on HepTh versus TGN at 51.70 and HGDHE at 57.98 on HepTh.
- Averaged over datasets, DHyperNodeTPP shows about 31.8% MRR improvement over the pairwise TGN model.
- On static directed hyperedge prediction, the proposed hyperedge predictor reaches MRR 61.3 on iAF1260b and 59.0 on iJO1366, above HyperSAGNN (60.1, 57.9) and CATSETMAT (35.0, 36.2).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use directed rather than undirected hyperedge modeling | DHyperNodeTPP vs HyperNodeTPP · HepTh and ML-Arxiv | MRR: 70% | Helped |
| Use directed rather than undirected hyperedge modeling | DHyperNodeTPP vs HyperNodeTPP · Eu-Email and Twitter | MRR: 75.95 ± 0.01 -> 68.05 ± 0.03; 84.47 ± 0.01 -> 82.12 ± 0.00 | Hurt |
| Use directed rather than undirected hyperedge modeling | DHyperNodeTPP vs HyperNodeTPP · Enron-Email and Bitcoin | MRR: 61.85 ± 0.01 -> 61.94 ± 0.01; 97.50 ± 0.09 -> 98.72 ± 0.05 | Helped |
| Use directed rather than undirected hyperedge modeling | DHyperNodeTPP vs HyperNodeTPP · Enron-Email, Eu-Email, Twitter, HepTh, ML-Arxiv, and Bitcoin | MAE: 4.15 ± 0.01 vs 4.18 ± 0.02; 12.23 ± 0.03 vs 12.22 ± 0.02; 1.18 ± 0.01 vs 1.20 ± 0.01; 1.25 ± 0.03 vs 1.20 ± 0.01; 1.25 ± | Mixed |
| Use DHyperNodeTPP instead of pairwise TGN | DHyperNodeTPP · Event type prediction | MRR: 31.8% | Helped |
| Use the proposed directed hyperedge predictor instead of HyperSAGNN | DHyperNodeTPP hyperedge predictor · iAF1260b, iJO1366, and USPTO | MRR: 3.7% | Helped |
| Use the proposed directed hyperedge predictor instead of CATSETMAT | DHyperNodeTPP hyperedge predictor · iAF1260b, iJO1366, and USPTO | MRR: 46.6% | Helped |
Surprising findings
- The undirected HyperNodeTPP outperforms the directed DHyperNodeTPP on several datasets (e.g., Enron-Email MRR 61.85 vs 61.94 is near-tie, Eu-Email 75.95 vs 68.05, Twitter 84.47 vs 82.12), so directionality does not uniformly help.
- On event time prediction, the directed and undirected models perform similarly, as events are modeled on nodes rather than hyperedges.
Limitations
- Undirected HyperNodeTPP matches or beats the directed model on several datasets, so the benefit of direction is dataset-dependent.
- Negative sampling is used to approximate the survival term of the node-event likelihood.
Experimental setup
Datasets: Enron-Email Eu-Email Twitter Hepth ML-Arxiv Bitcoin iAF1260b iJO1366 USPTO
Benchmarks: Enron-Email Eu-Email Twitter HepTh ML-Arxiv Bitcoin iAF1260b iJO1366 USPTO
Metrics: MRR MAE Recall AUC-macro MRR (static hyperedge prediction)
Baselines: GAT TGN HGDHE HGBDHE HyperNodeTPP HyperSAGNN CATSETMAT
Builds on: TGN HyperSAGNN CATSETMAT HGDHE GAT
Closest papers by meaning
- Deep Representation Learning for Forecasting Recursive and Multi-Relational Events in Temporal Networks (AAAI 2025 · similarity 0.92)
- Relational Learning in Pre-Trained Models: A Theory from Hypergraph Recovery Perspective (ICML 2024 · similarity 0.87)
- SPHINX: Structural Prediction using Hypergraph Inference Network (ICML 2025 · similarity 0.86)
- Training-Free Message Passing for Learning on Hypergraphs (ICLR 2025 · similarity 0.85)
- Improving Temporal Link Prediction via Temporal Walk Matrix Projection (NeurIPS 2024 · similarity 0.85)
- Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting (NeurIPS 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.