Explainable Artificial Intelligence for Multivariate Time Series Regression Models

Pedro André Gonçalves Zenário

Informações chave

Autores:

Pedro André Gonçalves Zenário (Pedro André Gonçalves Zenário)

Orientadores:

Luís Miguel Teixeira D'Avila Pinto da Silveira (Luís Miguel Teixeira D'Avila Pinto da Silveira); Ruxandra Georgeta Barbulescu

Publicado em

22/11/2023

Resumo

The rapid advancement of AI has revolutionized various industries by offering innovative solutions to intricate problems. However, this progress often results in complex, opaque models, making it challenging to comprehend their decision-making processes. This complexity is particularly concerning in vital sectors like medicine, where trust and transparency are paramount. To address this challenge, eXplainable AI has emerged as a pivotal research area. Understanding the underlying patterns and behaviors in data, especially within complex and dynamic systems, is essential for informed decision-making. This Thesis delves into XAI and Machine Learning interpretability, providing an extensive literature review that primarily focuses on elucidating the decision-making process of Deep Learning models in multivariate time-series regression. The study centers on the analysis of the responses of a roundworm’s nervous system (Caenorhabditis Elegans) under diverse stimuli. A GRU model was modified for this purpose, with adjustments in the last layer to align with the chosen explanation methods. The core of this research lies in a post-hoc explanation technique called SHAP, which not only identifies the features influencing the DL model’s decisions but also elucidates how features at different time points affect these decisions. The application of SHAP was validated through two experiments, involving limited input data during inference. These experiments corroborated the robustness and accuracy of SHAP explanations, demonstrating its potential to enhance the interpretability of DL models in MTS regression tasks. This study emphasizes the vital role of explainability in AI systems, enhancing trust and confidence in the decisions made by these complex algorithms.