Explainable AI (XAI) refers to methods to make the operations of AI systems more understandable to humans. Post-hoc methods are a subset of XAI techniques used after a model has made its predictions to elucidate the decision-making process. These methods aim to interpret and explain advanced models' often complex and opaque functioning, such as neural networks. These methods include feature importance ranking, which identifies which input features most influenced the predictions; partial dependence plots, which illustrate relationships between features and predictions; surrogate models, such as decision trees, that approximate the behavior of complex models; LIME, which creates interpretable, local surrogate models around specific predictions to explain individual decisions; and SHAP, which assigns each feature an importance value for a particular prediction based on cooperative game theory concepts. Post-hoc methods help improve transparency, allow easier debugging, and are an emerging field within machine learning.
This Master's project will examine the applicability of XAI techniques to support and enhance software development, making software behavior more transparent and understandable for developers and stakeholders. We will investigate how XAI can be applied in areas such as behavior explanation, misbehavior detection, testing enhancement, conformance verification, and specification generation.