Evidential Temporal-aware Graph-based Social Event Detection via Dempster-Shafer Theory

Jiaqian Ren, Lei Jiang, Hao Peng, Zhiwei Liu, Jia Wu, Philip S. Yu

Social media platforms serve as a vital pulse for real-time global events. However, identifying specific events within the deluge of noisy, unstructured data remains a significant challenge for automated systems. While Graph Neural Networks (GNNs) have shown promise by modeling relationships between tweets, they often struggle with the varying reliability of data "views" (e.g., hashtags vs. user mentions) and the critical role of temporal proximity. This paper introduces the Evidential Temporal-aware Graph Neural Network (ETGNN), a framework designed to address these limitations by integrating temporal decay and Dempster-Shafer Theory to quantify and manage uncertainty.

The Limitations of Current GNN Approaches

Existing GNN-based event detection models typically construct graphs where nodes represent tweets and edges represent shared features like hashtags, entities, or users. However, these models face two primary hurdles:

  1. Heterogeneous View Reliability: Not all shared features are equally informative. A shared niche hashtag is a strong signal of a specific event, whereas a shared common user or a generic hashtag may be mere noise. Standard models often fail to distinguish between these "views" when one is untrustworthy.
  2. Temporal Dynamics: Events are inherently time-bound. The relevance between two tweets decays as the time interval between them increases. Many models treat time as a static feature or ignore it entirely, failing to capture the "bursty" nature of social events.

The ETGNN Architecture

ETGNN introduces a multi-view approach that explicitly models both time and uncertainty through three core innovations.

1. Temporal-aware GNN Aggregator

Unlike standard aggregators that treat neighbors equally, ETGNN employs an adaptive time exponential decay formula. The decay rate, $\lambda$, is not a fixed hyperparameter but is learnable based on the target node’s features. This allows the model to adapt to the varying "life cycles" of different event types—recognizing that some events are short-lived flashes while others persist for days.

2. Evidential Deep Learning (EDL)

To move beyond the potential overconfidence of standard Softmax layers, ETGNN utilizes Evidential Deep Learning. By placing a Dirichlet distribution over class probabilities, the model can quantify the uncertainty associated with each view (hashtag, entity, and user). This allows the system to effectively "know what it doesn't know."

3. Dempster-Shafer Theory (DST) Fusion

The core of ETGNN’s robustness lies in its fusion mechanism. Using Dempster-Shafer Theory, the model combines the evidence from different views. If the "hashtag view" is highly uncertain due to noise or hijacking, DST allows the model to dynamically prioritize evidence from more certain views, such as "entities" or "users." A consistency constraint is also applied to ensure that representations across different views maintain a similar internal structure.

Experimental Evaluation

The researchers evaluated ETGNN against state-of-the-art baselines (including KPGNN and MVGAN) using three diverse real-world datasets: Events2012 (68,841 English tweets), Kawarith (9,070 Arabic crisis tweets), and CrisisLexT26 (6,733 English crisis tweets).

  • Superior Performance: On the Events2012 dataset, ETGNN achieved an F1 score of 0.7565, outperforming the next best model, MVGAN (0.7358).
  • Resilience to Noise: In stress tests where Gaussian noise was introduced to the hashtag view, ETGNN maintained stable performance. In contrast, traditional fusion methods like simple averaging or standard attention mechanisms saw significant drops in accuracy.
  • Uncertainty Quantification: The model successfully assigned higher uncertainty scores to samples with higher noise levels, validating the effectiveness of the EDL component in identifying unreliable data.

Conclusion and Future Work

ETGNN represents a significant step forward in making social event detection more robust and interpretable. By treating time as a learnable decay and using evidence theory to manage multi-view uncertainty, it effectively filters noise from signal. While the model currently faces some computational overhead and challenges with highly conflicting evidence, future research will focus on optimizing the fusion process and scaling the system to handle massive, high-velocity social media streams in real-time.

Summary updated

BibTeX
@inproceedings{ren2022evidential,
  author = {Jiaqian Ren and Lei Jiang and Hao Peng and Zhiwei Liu and Jia Wu and Philip S. Yu},
  title = {{Evidential Temporal-aware Graph-based Social Event Detection via Dempster-Shafer Theory}},
  booktitle = {ICWS},
  year = {2022},
  url = {https://openreview.net/forum?id=MqrSYjtB5I},
}
Figure