• PhD Student:
  • Steve Bellart
  • Funding : ANR
  • PhD defended on :
  • Dec 18, 2023

Artificial Intelligence (AI) empowers machines to make autonomous decisions. Over time, empha- sis has been placed on machines’ ability to learn and make decisions independently, utilising machine learning methods. However, a significant challenge posed by the derived systems is their black box na- ture, making their decision-making process difficult to understand. This opacity becomes particularly problematic in fields such as medicine, finance, recruitment, and the judicial system, where unexplained decisions can have profound implications. Recognizing the importance of transparency, the European Union (EU) has implemented regulations like the General Data Protection Regulation (GDPR) and pro- posed new guidelines in 2021 to ensure ethical and transparent use of AI. The demand for clarity not only fosters trust but also aligns with regulatory requirements.

In this thesis, we examine the robust explanation of predictions made by ensemble tree models, particularly random forests and boosted trees. Although these models are based on decision trees, which are naturally considered as interpretable, their predictions can be challenging to explain. In the current AI landscape, trust, ethics, interpretability, and explainability are of paramount importance. We propose tailored solutions for explaining these two types of models. For random forests, we define a new notion of formal explanation that is both concise and computable in polynomial time. For boosted tree models, such as those trained with libraries like LGBM and XGBoost, we expose a methodology for explaining the results obtained in regression problems, seeking to elucidate why a prediction falls within a certain interval. Finally, we introduce notions of personalized explanations leveraging users’ knowledge bases and preferences.