همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

I2KAN: Interpretable Interaction-aware Kolmogorov-Arnold Networks for Time-Series Anomaly Detection

نویسندگان :

Amirhossein Sadr¹ Mahboobeh Riahi-Madvar²

1- Shahid Beheshti University 2- Shahid Beheshti University

کلمات کلیدی :

Time Series Anomaly Detection،Kolmogorov-Arnold Networks،Interpretability،Multivariate Interaction،Deep Learning

چکیده :

Time Series Anomaly Detection (TSAD) is critical for the reliability of modern IT infrastructure, cloud services, and industrial systems. While deep learning models have advanced forecasting-based TSAD, they often overfit to local fluctuations and lack interpretability, limiting their trustworthiness in safety-critical applications. This paper introduces I²KAN, a novel framework that leverages Kolmogorov-Arnold Networks (KAN) to address these dual challenges. I²KAN enhances KAN’s architecture with two key strategies: an interaction-aware feature weighting mechanism that dynamically quantifies and incorporates the multivariate dependencies within time series data, and an interpretability pipeline that provides transparent rationales for its predictions. By moving beyond the black-box nature of traditional deep models, I²KAN not only achieves high detection accuracy but also offers crucial insights into the underlying anomaly patterns. Extensive experiments on the ECG5000 and UCR benchmarks demonstrate that I²KAN achieves state-of-the-art performance, outperforming MLP, KAN, and KAN-AD models in Accuracy, ROC-AUC, and F1 score. Furthermore, we employ SHAP analysis to validate and visualize the model's decision-making process, reinforcing its credibility for real-world deployment.