Local Model-Agnostic Methods for Model Explainability

Most machine and deep learning models used today are inherently opaque: their internal mechanisms remain hidden and are difficult for humans to interpret, often leading them to be referred to as "black-box systems". As these models are seeing an increasing deployment in critical domains, it is crucial that users, decision-makers and stakeholders understand the rationale behind their outputs. In order to address this concern, Explainable Artificial Intelligence (XAI) has emerged as a field with the purpose of providing human-understandable insights into the workings of models, in order to enhance their transparency and trust. This thesis work aims to introduce some of the most relevant post hoc techniques for model explainability, specifically focusing on local model-agnostic methods, including LIME (Local Interpretable Model-agnostic Explanations), Anchors and SHAP (SHapley Additive exPlanations). These techniques produce explanations on a per-instance basis, as they are relative to a specific prediction for a particular input, and they can be applied to any model type, regardless of the underlying architecture. An experimental phase was conducted using the Amazon Review dataset for sentiment analysis classification, using LIME and SHAP to explain text instances based on predictions obtained by three distinct models of increasing complexity: SVM, LSTM and the pre-trained Transformer model RoBERTa. The results highlighted that these methods provide meaningful information, particularly in identifying words or phrases that significantly influenced the final models' text classifications, enabling a deeper understanding of their internal workings.