| Title |
Explainable Graph Neural Network-based Time Series Anomaly Detection |
| Authors |
이준호(Junho Lee) ; 박예인(Ye-In Park) ; 조민지(Minji Cho) ; 강석주(Suk-Ju Kang) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.1.83 |
| Keywords |
Anomaly detection; Graph deviation network; Time series; Wavelet transform |
| Abstract |
Time series anomaly detection aims to identify unusual patterns that emerge over time in given data. Several studies have been increasingly used the explainable artificial intelligence in various tasks to enhance the explainability of the anomaly detection. However, traditional research has not effectively considered features that differ significantly from existing feature. In order to solve this problem, we introduce a novel framework that applies shapley additive explanations to understand the importance of input features on the model output. Specifically, dissimilar features have the negative effect on learning a given model. Therefore, we selected features based on shap values according to their importance and used them in training. Additionally, we present a method that transforms the given data into a graph structure and applies wavelet transformation to highlight anomalies. This framework not only makes it easier for the model to identify anomalies but also enhances the explainability of the model. Evaluation on widely-used benchmark datasets shows our method outperforms the previous methods. Our framework achieves 10% F1 score improvements compared to the previous methods. |