Abstract:Point process models are widely used for continuous asynchronous event data, where each data point includes time and additional information called "marks", which can be locations, nodes, or event types. This paper presents a novel point process model for discrete event data over graphs, where the event interaction occurs within a latent graph structure. Our model builds upon Hawkes's classic influence kernel-based formulation in the original self-exciting point processes work to capture the influence of historical events on future events' occurrence. The key idea is to represent the influence kernel by Graph Neural Networks (GNN) to capture the underlying graph structure while harvesting the strong representation power of GNNs. Compared with prior works focusing on directly modeling the conditional intensity function using neural networks, our kernel presentation herds the repeated event influence patterns more effectively by combining statistical and deep models, achieving better model estimation/learning efficiency and superior predictive performance. Our work significantly extends the existing deep spatio-temporal kernel for point process data, which is inapplicable to our setting due to the fundamental difference in the nature of the observation space being Euclidean rather than a graph. We present comprehensive experiments on synthetic and real-world data to show the superior performance of the proposed approach against the state-of-the-art in predicting future events and uncovering the relational structure among data.
What problem does this paper attempt to address?
The main problem this paper attempts to address is: how to model point processes in discrete event data on graph structures to capture the complex dynamic influences between events. Specifically, the paper proposes a novel point process model based on Graph Neural Networks (GNN) to represent discrete event data on graph structures. This model uses GNN to represent the influence kernel function in the classical self-exciting point process, thereby better capturing the impact of historical events on future events. It also effectively combines the advantages of statistical models and deep learning models, improving the learning efficiency and predictive performance of the model.
### Main Contributions
1. **Explicit Modeling of Influence Kernels**: Unlike traditional intensity function-based models, this method explicitly models the influence kernels in point processes using GNN, allowing for more complex inter-event contributions, including non-stationary, multi-hop excitation, and inhibition effects. Additionally, deep graph kernels can capture dynamic influences, providing clear information on the relational structure in graph point processes.
2. **Local Graph Filter Basis**: The deep kernel surpasses simple distance-based influences through local graph filter bases, making it suitable for non-spatial graphs (such as traffic or social networks). It can also incorporate more types of GNN models within the framework, enhancing practical application flexibility.
3. **Experimental Validation**: Comprehensive experiments demonstrate the advantages of deep kernel modeling with latent graph structures on both synthetic and real datasets. This method is applicable to various point process data scenarios, including events generated by infrastructure, climate, and social phenomena.
### Background
- **Self-Exciting Point Processes**: Self-exciting point process models describe the occurrence of timestamped discrete events that depend on observed history. The classic Hawkes process models self-exciting effects in history additively.
- **Marked Point Processes**: Each event is associated with an additional attribute (mark), which can be continuous or categorical, such as event location or type. The conditional intensity function considers the influence of time and marks.
- **Graph Convolution**: Graph convolution extends convolution strategies to graph structures, divided into spectral and spatial models. Spectral graph convolution is based on the eigen decomposition of the graph Laplacian matrix, while spatial graph convolution propagates information along edges.
### Method
- **Deep Temporal Graph Kernels**: Represent multidimensional influence kernels through kernel singular value decomposition (SVD), capturing heterogeneous temporal dependencies between past and future events.
- **Graph Basis Kernels and Local Graph Filters**: Utilize local graph filters in graph convolution to extract information patterns in graph-structured data, representing local and global patterns of complex node influences.
- **Model Estimation**: Use negative log-likelihood (NLL) and least squares (LS) loss functions for model estimation, ensuring the non-negativity of the conditional intensity function.
### Experiments
- **Synthetic Data**: Evaluate the model's effectiveness on 5 synthetic datasets generated by point processes with different kernels and latent graph structures.
- **Known Graph Topology**: Demonstrate GraDK's superior performance in modeling point process data using observed latent graph structures through synthetic datasets.
- **Dynamic Influence**: Figure 2 shows the dynamic influences learned by GraDK on a 50-node graph.
- **Event Dependency**: The second row of Figure 1 shows the event dependencies learned in test sequences generated on a 16-node graph.
- **Predictive Ability**: Further demonstrate GraDK's superior predictive performance through comprehensive evaluations.
In summary, this paper significantly extends existing spatiotemporal point process models by introducing GNN-based deep graph kernels, making them suitable for discrete event data on graph structures and improving the model's learning efficiency and predictive performance.