Learning Neural Jump Stochastic Differential Equations with Latent Graph for Multivariate Temporal Point Processes

Yuchen Wang, Dongpeng Hou, Chao Gao, Xianghua LiIJCAI 2025 · DOI

Presented at IJCAI-25, this research introduces LNJSDE (Learning Neural Jump Stochastic Differential Equations with Latent Graph), a framework designed to predict the timing and type of events in complex sequences. By integrating the continuous modeling capabilities of Stochastic Differential Equations (SDEs) with a learned latent graph of event relationships, the authors achieve state-of-the-art performance in event forecasting and model interpretability.

Abstract

LNJSDE models multivariate event sequences by utilizing multi-dimensional SDEs to track hidden states and a learned latent graph to govern how different event types influence one another. This approach outperforms existing neural point processes on real-world datasets while providing a clear, interpretable map of inter-event dependencies.

The Challenge of Multivariate Temporal Point Processes

Many real-world systems generate asynchronous event sequences, such as stock market trades, social media interactions, or seismic activity. Researchers use Multivariate Temporal Point Processes (MTPPs) to model these systems. The primary challenge lies in simultaneously capturing two dynamics:

  1. Continuous Dynamics: How the "intensity" (likelihood) of an event evolves over time between actual occurrences.
  2. Inter-event Relationships: How an event of one type (e.g., a "like") affects the probability of another event type (e.g., a "share").

Existing methods often simplify these relationships or rely on rigid mathematical assumptions like the Hawkes process. While neural versions offer more flexibility, they frequently struggle to model the underlying "graph" of influences, leading to limited interpretability and predictive power.

Methodology: Jump-Diffusion with Latent Graphs

LNJSDE proposes a hybrid approach that treats event sequences as a Jump-Diffusion process. The architecture consists of several key components:

  • Multi-dimensional Neural SDEs: Instead of 1D variables, the model uses multi-dimensional hidden states that evolve continuously according to an SDE, accounting for both predictable trends (drift) and random noise (diffusion).
  • Encoder-Free Latent Graph: The model learns a "latent graph" as a set of trainable parameters ($\Psi$). This decouples the graph structure from specific hidden states, enhancing training stability.
  • The Jump Process: When an event occurs, it triggers a "jump" in the hidden states of related event types. The magnitude and direction of this jump are guided by the learned latent graph.
The LNJSDE framework: A latent graph $\Psi$ is transformed into an adjacency matrix $A$ via Softmax, which then governs the jump process affecting the intensity functions of various event types.

Experimental Results

The model was evaluated on five real-world benchmarks: NYMVC (vehicle collisions), MathOF and AskUbuntu (Stack Exchange interactions), Taobao (e-commerce behavior), and Taxi (trip records).

Predictive Performance

LNJSDE consistently outperformed eight state-of-the-art baselines, including Transformer-based and ODE-based models. It achieved the lowest Negative Log-Likelihood (NLL) across nearly all datasets. For instance, in the AskUbuntu dataset, LNJSDE reached an accuracy of 0.391, significantly higher than the next best model, NJDTPP (0.358).

Interpretability and Graph Recovery

A standout feature of LNJSDE is its ability to recover the underlying "Granger causality" between event types. On synthetic data with a known influence structure, LNJSDE achieved an AUC of 0.829 for graph recovery, far surpassing the 0.603 achieved by VAETPP.

ROC curve for graph recovery on synthetic data, demonstrating that LNJSDE (Our SDE) significantly outperforms the baseline VAETPP in identifying true inter-event relationships.

Key Insights

  • Efficiency: By passing messages between nodes only at the moment of an event (the jump), the model avoids the high computational cost of continuous message passing.
  • Linear Noise Robustness: The authors found that a Linear Noise SDE (LNSDE) variant was particularly effective at preventing overfitting to noise in time intervals.

Limitations and Future Work

While powerful, the authors identify scalability to extremely large-scale graphs (e.g., social networks with thousands of types) as a remaining challenge. Future work will also explore "non-stationary" networks where the graph structure itself evolves over time.

Summary updated

BibTeX
@inproceedings{wang2025learning,
  author = {Yuchen Wang and Dongpeng Hou and Chao Gao and Xianghua Li},
  title = {{Learning Neural Jump Stochastic Differential Equations with Latent Graph for Multivariate Temporal Point Processes}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/383},
}

Loading related content...

Figure