Neural Temporal Point Processes for Forecasting Directional Relations in Evolving Hypergraphs

Tony Gracious, Arman Gupta, Ambedkar DukkipatiAAAI 2025 · DOI

TL;DR

This paper introduces DHyperNodeTPP, a novel framework for predicting directional, higher-order relationships in dynamic networks. It tackles the challenge of forecasting evolving hypergraphs by breaking down the problem into sequential stages: node event prediction, candidate hyperedge generation, and directed hyperedge prediction, significantly outperforming existing methods.

What Problem Does This Paper Address?

Real-world relationships are often complex: they involve more than two entities (higher-order), have a clear direction (e.g., sender/receiver), and evolve over time. Existing research often tackles these aspects in isolation, leading to models that can't fully capture the richness of these dynamic, higher-order, directional networks. For example, predicting future interactions in a network like Bitcoin transactions, where multiple addresses send funds to multiple other addresses at specific times, is a significant challenge. The paper aims to provide a unified solution for forecasting these intricate relationships in evolving hypergraphs.

Methods

The core of the proposed solution is the DHyperNodeTPP model, which employs a sequential generative approach with three key modules:

  • Node Event Model: This module uses a Temporal Point Process (TPP) to predict when and which nodes are likely to be involved in future events. This is a departure from traditional methods that might predict hyperedge events directly, reducing the search space.
  • Candidate Generation: Based on the predicted node events, this module generates potential hyperedges. It predicts projected adjacency matrices and hyperedge sizes for the involved nodes, creating a set of likely candidates.
  • Hyperedge Predictor: This module then filters the candidate hyperedges to identify the true, directed hyperedges. It leverages a novel architecture that considers three types of information within a directed hyperedge: self-connections within the left and right sets of nodes, and cross-connections between them.

A key innovation is the use of a Memory Module and temporal message-passing to enable batch processing, which is crucial for scalability in large datasets. This contrasts with previous sequential models that processed data sample-by-sample.

Datasets

The paper introduces five newly compiled real-world temporal-directed hypergraph datasets for evaluation. These datasets are derived from open-source data. The paper also lists several existing datasets used for comparison, including static and temporal directed hypergraphs.

Key Insights and Ideas

  • Decomposition is Key: Breaking down the complex problem of forecasting directed hyperedges into sequential stages (node events -> candidate generation -> prediction) significantly reduces computational complexity and improves accuracy.
  • Direction Matters: Explicitly modeling the directionality of hyperedges, not just their existence, leads to better performance, especially in tasks like adjacency forecasting and event type prediction.
  • Batch Processing for Scalability: The integration of a Memory Module and temporal message-passing is critical for making these models scalable to large, real-world datasets, a limitation of prior sequential approaches.
  • Higher-Order Structure is Crucial: Models that can effectively capture the relationships within hyperedges (not just pairwise) demonstrate superior performance, particularly for larger hyperedges.

Key Results

The proposed DHyperNodeTPP model achieved significant performance gains:

  • Event Type Prediction: An average improvement of 32% and 41% compared to state-of-the-art pairwise and hyperedge event forecasting models, respectively.
  • Directed vs. Undirected: DHyperNodeTPP showed an average increase of 70% in MRR for event type prediction on citation network datasets compared to its undirected counterpart, HyperNodeTPP.
  • Hyperedge Prediction: The proposed hyperedge predictor outperformed previous works on static directed hyperedge prediction, achieving an average improvement of 3.7% over undirected models and 46.6% over bipartite models in MRR.
  • Temporal Models Outperform Static: Temporal models, including DHyperNodeTPP, consistently outperformed static models like GAT, highlighting the importance of temporal dynamics.

Limitations and Future Work

  • Scalability: While the paper addresses scalability with batch processing, the inherent complexity of hypergraphs remains a challenge for extremely large-scale scenarios.
  • Dataset Creation: The paper mentions the creation of five new datasets, suggesting that further dataset development and standardization could benefit the field.
  • Future Directions: The authors suggest that future work could explore more sophisticated architectures for hyperedge prediction and further investigate the impact of directionality across various real-world applications. The paper also notes that the model's performance on event time prediction is similar to its undirected counterpart, suggesting that further research might be needed to fully leverage directionality for temporal forecasting.

Summary updated

BibTeX
@inproceedings{gracious2025neural,
  author = {Tony Gracious and Arman Gupta and Ambedkar Dukkipati},
  title = {{Neural Temporal Point Processes for Forecasting Directional Relations in Evolving Hypergraphs}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i16.33856},
}
Figure