Interpreting Long Short-Term Memory Autoencoders: A Study on Explainable AI Using LIME

Sammanfattning

As machine learning models become increasingly prevalent in financial applications, the need for transparency and interpretability has grown. This thesis investigates the application of Explainable Artificial Intelligence (XAI) in the context of anomaly detection within transactional data at Handelsbanken. Specifically, Long Short-Term Memory (LSTM) autoencoders are used to detect anomalies in sequential trading data, followed by the application of the Local Interpretable Model-agnostic Explanations (LIME) method to interpret the models' behaviour. The study highlights the effectiveness of LSTM networks in capturing temporal dependencies and reconstructing legitimate transaction sequences. To bridge the gap between predictive performance and interpretability, reconstruction errors from the autoencoder are leveraged as labels for the LIME explanations. Through single instance analysis, aggregated importance measures, and heatmaps, the study demonstrates how LIME can provide valuable insights into feature contributions at local and global levels. The results indicate that LIME can be a practical tool for post hoc model interpretation in financial anomaly detection, offering potential benefits for improved transparency and model trustworthiness within the banking sector.