Exploring the accuracy-size trade-off in extended argumentative decision graphs using SHAP or LIME feature selection

Structured argumentation frameworks provide a promising approach to enhancing the explainability of machine learning (ML) models. This study employs graph-like models learnt from data that can be used for classification tasks, also known as eXtended Argumentative Decision Graphs (xADGs). We leverage SHAP or LIME values to reduce inference model size while analysing the trade-off with predictive performance. By selecting the most influential features, an acceptable reduction in accuracy is expected. At the same time, a notable gain in explainability is also anticipated, as this selection can likely minimise the number of arguments and attacks related to them. Several thresholds for SHAP and LIME values were evaluated to identify the most influential features, and the corresponding impacts on the related xADGs were subsequently analysed. In brief, this study further explores enhancements to xADGs, aiming to realise their potential to develop accurate and concise structured argumentative models.