EXPLAINABLE AI FOR TIME SERIESANOMALY DETECTION

Sammanfattning

The increasing use of complex Artificial Intelligence (AI) models, particularly in time seriesanomaly detection across various domains like finance, industry, and cybersecurity, has highlighteda critical need for transparency and interpretability. Many current AI systems operate as ’blackboxes’, making it difficult to understand their decision-making processes, which hinders trust andadoption, especially in critical applications. Existing anomaly detection tools often lack essentialfeatures such as user-friendly interfaces, modularity, and, most importantly, the ability to explainwhy a data point is classified as anomalous.This thesis addresses these challenges by introducing the Explainable Anomaly ClassificationTool (EXACT), a novel, modular software tool, integrating time series anomaly detection with XAItechniques. EXACT facilitates data handling, anomaly injection, model training, and explanationgeneration. We evaluate Extreme Gradient Boosting (XGBoost), Decision Tree (DT), and LongShort-Term Memory with Autoencoder (LSTM-AE) models alongside SHAP, LIME, and DiCEmethods across four benchmark datasets. Machine Learning performance is assessed using metricslike F1-score and recall, while XAI methods are evaluated via Normalised Discounted CumulativeGain (NDCG) for feature importance and counterfactual analysis.Key findings show XGBoost offers balanced performance, DT is fastest, and LSTM-AE is com-putationally intensive but captures temporal aspects. SHAP provides robust feature rankings (highNDCG) but can be slow; LIME offers faster but sometimes less reliable explanations; and DiCE gen-erates valuable "what-if" counterfactuals. The study highlights crucial trade-offs between accuracy,interpretability, and computational cost. The EXACT tool and our empirical insights contributeto advancing trustworthy AI by providing a practical tool and guidance for applying and evaluatingexplainable anomaly detection, making these techniques more accessible and understandable.