[2308.16245] Calibrated Explanations for Regression

View a PDF of the paper titled Calibrated Explanations for Regression, by Tuwe L”ofstr”om and 4 different authors

View PDF
HTML (experimental)

Summary:Synthetic Intelligence (AI) is commonly an integral a part of fashionable resolution help methods. The most effective-performing predictive fashions utilized in AI-based resolution help methods lack transparency. Explainable Synthetic Intelligence (XAI) goals to create AI methods that may clarify their rationale to human customers. Native explanations in XAI can present details about the causes of particular person predictions when it comes to characteristic significance. Nonetheless, a essential downside of current native rationalization strategies is their incapacity to quantify the uncertainty related to a characteristic’s significance. This paper introduces an extension of a characteristic significance rationalization technique, Calibrated Explanations, beforehand solely supporting classification, with help for normal regression and probabilistic regression, i.e., the chance that the goal is above an arbitrary threshold. The extension for regression retains all the advantages of Calibrated Explanations, comparable to calibration of the prediction from the underlying mannequin with confidence intervals, uncertainty quantification of characteristic significance, and permits each factual and counterfactual explanations. Calibrated Explanations for normal regression offers quick, dependable, secure, and strong explanations. Calibrated Explanations for probabilistic regression offers a completely new approach of making probabilistic explanations from any strange regression mannequin, permitting dynamic choice of thresholds. The tactic is mannequin agnostic with simply understood conditional guidelines. An implementation in Python is freely accessible on GitHub and for set up utilizing each pip and conda, making the outcomes on this paper simply replicable.