Graph Evidential Learning for Anomaly Detection

Chunyu Wei1     Wenji Hu1     Xingjia Hao2     Yunhai Wang1     Yueguo Chen1     Bing Bai3     Fei Wang4    

1Renmin University of China     2Guangxi University     3Microsoft MAI
4Weill Cornell Medicine    

Accepted by ACM SIGKDD Conference on Knowledge Discovery and Data Mining

Figure 1: Introducing Uncertainty for anomaly detection.Relying on reconstruction error, nodes 3 and 4 in the graph are misclassified without introducing uncertainty.


Abstract:

While existing visualization libraries enable the reuse, extension, and combination of static visualizations, achieving the same for interactions remains nearly impossible. Therefore, we contribute an interaction model and its implementation to achieve this goal. Our model enables the creation of interactions that support direct manipulation, enforce software modularity by clearly separating visualizations from interactions, and ensure compatibility with existing visualization systems. Interaction management is achieved through an instrument that receives events from the view, dispatches these events to graphical layers containing objects, and then triggers actions. We present a JavaScript prototype implementation of our model called Libra, enabling the specification of interactions for visualizations created by different libraries. We demonstrate the effectiveness of Libra by describing and generating a wide range of existing interaction techniques. We evaluate Libra.js through diverse examples, a metric-based notation comparison, and a performance benchmark analysis.

Source Code: https://github.com/wuanjunruc/GEL




Figures:





Figure 2: The overview of the GEL framework. GEL models high-order evidential distribution to calculate feature uncertainty and topology uncertainty during the graph reconstruction process, leveraging these uncertainties for anomaly detection.



Figure 3: Heatmaps of average normalized anomaly scores on the Cora dataset. Rows correspond to the class omitted during training (anomalous class), columns represent the class labels during testing.



Figure 4: Impact of hidden layer dimension. GEL’s performance on the Weibo and Disney datasets as the latent dimension 𝑑′ increases. We observe that performance improves with larger 𝑑′, enabling the model to capture more reconstruction evidence and leading to a more accurate high-order evidential distribution, thus enhancing uncertainty quantification.



Figure 5: Impact of removing Different Modality. GEL achieves optimal performance when uncertainties from both modalities are considered; removing any modality results in a noticeable performance decline. The performance degradation is more pronounced when the feature modality is omitted, indicating that uncertainty in node attributes plays a critical role in anomaly detection. This suggests that node attributes provide richer information, enhancing the model’s capability to detect anomalies.



Figure 6: Changes of AUC and Recall@K as the level of data disturbance increases. GEL’s robustness arises from its construction of high-order evidential and reconstruction distributions by capturing evidence from multiple dimensions to quantify uncertainty. This enables GEL to better handle missing or noisy data, maintaining relatively stable performance even under significant perturbations.



Materials:





1
Paper (1.01M)

Acknowledgements:

This research was supported by the National Key R&D Program of China (No. 2023YFC3304701) and in part by the Young Elite Scientists Sponsorship Program by CAST under contract No. 2022QNRC001. It was also supported by Big Data and Responsible Artificial Intelligence for National Governance, Renmin University of China.