Abstract
Using black box machine learning models (e.g., Deep Neural Networks) in high-stakes domains such as healthcare, criminal justice and real-time systems can cause serious problems due to their complexity and poor interpretability. Moreover, model selection with interpretability in addition to accuracy is one of emerging research areas with lack of model agnostic and quantitative interpretability metrics. In this work, we adopt a quantitative interpretability metric, and then, introduce a trade-offs methodology between accuracy and interpretability, which can be demonstrated by increasing interpretability of ML models while allowing accuracy to drop up to given thresholds. In our experimental results, interpretability in terms of simulatability operation count (SOC) is improved up to 76.2% with minimal 2.3% accuracy drop in a SVR estimator of the Auto MPG dataset (up to 64.3% with minimal 1.9% accuracy drop in the Forest Fire dataset of an MLP estimator).