As machine learning (ML) models become increasingly complex, ensuring their interpretability has become a focal point in recent research. A study by Schmidt and Biessmann (2019) introduced quantitative measures to assess the quality of interpretability methods and the trust in ML decisions. They proposed evaluating the intuitive understanding of algorithmic decisions using the information transfer rate at which humans replicate ML model predictions. Their empirical experiments demonstrated that providing explanations significantly improved productivity in annotation tasks, underscoring the value of interpretability in ML-assisted human decision-making. arxiv.org
However, the pursuit of interpretability is not without challenges. Luo et al. (2024) highlighted privacy risks associated with Shapley value-based interpretability methods, which are widely adopted by leading ML as a Service providers. They demonstrated that feature inference attacks could reconstruct private model inputs based on their Shapley value explanations, emphasizing the need for privacy-preserving interpretability methods. arxiv.org Additionally, Bassan et al. (2024) examined the computational complexity of local and global interpretability, revealing that the difficulty of computing explanations varies across different model types, such as linear models, decision trees, and neural networks. These findings suggest that a nuanced approach is necessary when developing interpretability methods tailored to specific ML models. arxiv.org