DyGCL: Dynamic Graph Contrastive Learning For Event Prediction

TL;DR

DyGCL tackles event prediction from text-derived dynamic graphs by learning both local node-level dynamics and global graph-level evolution through two dedicated encoders. A contrastive objective aligns the two views, and their fused representation fed to an MLP predicts future events. The approach achieves state-of-the-art results on six real-world datasets and demonstrates the complementary value of local and global graph information, with extensive ablations highlighting the contribution of each component. This dual-view, contrastive framework offers a principled way to capture temporal patterns in textual event contexts and can be extended to diverse event-prediction tasks.

Abstract

Predicting events such as political protests, flu epidemics, and criminal activities is crucial to proactively taking necessary measures and implementing required responses to address emerging challenges. Capturing contextual information from textual data for event forecasting poses significant challenges due to the intricate structure of the documents and the evolving nature of events. Recently, dynamic Graph Neural Networks (GNNs) have been introduced to capture the dynamic patterns of input text graphs. However, these models only utilize node-level representation, causing the loss of the global information from graph-level representation. On the other hand, both node-level and graph-level representations are essential for effective event prediction as node-level representation gives insight into the local structure, and the graph-level representation provides an understanding of the global structure of the temporal graph. To address these challenges, in this paper, we propose a Dynamic Graph Contrastive Learning (DyGCL) method for event prediction. Our model DyGCL employs a local view encoder to learn the evolving node representations, which effectively captures the local dynamic structure of input graphs. Additionally, it harnesses a global view encoder to perceive the hierarchical dynamic graph representation of the input graphs. Then we update the graph representations from both encoders using contrastive learning. In the final stage, DyGCL combines both representations using an attention mechanism and optimizes its capability to predict future events. Our extensive experiment demonstrates that our proposed method outperforms the baseline methods for event prediction on six real-world datasets.

Figures (6)

• Figure 1: An overview of Dynamic Graph Contrastive Learning, (DyGCL) architecture. We feed input graphs into the Local View Encoder to learn the dynamic node representations. In the end, node representations are pooled into a graph-level representation using a pooling layer. We also feed input graphs into the Global View Encoder to learn dynamic graph representation. Use Contrastive learning to maximize the similarity between two graph representations. Finally, representations from Local View Encoder and Global View Encoder are combined by an MLP layer and feed the representation to the predictor for event prediction.
• Figure 2: Detailed Event prediction results of $\texttt{DyGCL}_{sup}$ model without contrastive learning for the different number of historic days.
• Figure 3: Detailed Event prediction results of DyGCL model for the different number of historic days.
• Figure 4: Detailed event prediction results of $\texttt{DyGCL}_{sup}$ model without contrastive loss for the different number of lead days.
• Figure 5: Detailed event prediction results of the DyGCL model for the different number of lead days.

Theorems & Definitions (2)

• Definition 1: Dynamic Graph
• Definition 2: Event Prediction