In this study, two groups of respondents have evaluated explanations generated from an instance-based explanation method called WITE (Weighted Instance-based Text Explanations). One group consisted of 24 non-experts who answered a web survey about the words characterising the concepts of the classes and the other group consisted of three senior researchers and three respondents from a media house in Sweden who answered a questionnaire with open questions. The data used originates from one of the researchers’ project on media consumption in Sweden. The results from the non-experts indicate that WITE identified many words that corresponded to the human understanding but also included some insignificant or contrary words as important. In the results from the expert evaluation, there were indications that there is a risk that the explanations could persuade the users of the correctness of a prediction, even if it is incorrect. Consequently, the study indicates that an explanation method could be seen as a new actor which is able to persuade and interact with the humans and cause a change in the results of the classification of a text.

The evaluation of explanation methods has become a significant issue in explainable artificial intelligence (XAI) due to the recent surge of opaque AI models in decision support systems (DSS). Explanations are essential for bias detection and control of uncertainty since most accurate AI models are opaque with low transparency and comprehensibility. There are numerous criteria to choose from when evaluating explanation method quality. However, since existing criteria focus on evaluating single explanation methods, it is not obvious how to compare the quality of different methods.

Evaluating the efficiency of human-AI interactions is challenging, including subjective and objective quality aspects. With the focus on the human experience of the explanations, evaluations of explanation methods have become mostly subjective, making comparative evaluations almost impossible and highly linked to the individual user. However, it is commonly agreed that one aspect of explanation quality is how effectively the user can detect if the predictions are trustworthy and correct, i.e., if the explanations can increase the user's appropriate trust in the model. This paper starts with the definitions of appropriate trust from the literature. It compares the definitions with model performance evaluation, showing the strong similarities between appropriate trust and model performance evaluation. The paper's main contribution is a novel approach to evaluating appropriate trust by taking advantage of the likenesses between definitions. The paper offers several straightforward evaluation methods for different aspects of user performance, including suggesting a method for measuring uncertainty and appropriate trust in regression.

Predictive models used in Decision Support Systems (DSS) are often requested to explain the reasoning to users. Explanations of instances consist of two parts; the predicted label with an associated certainty and a set of weights, one per feature, describing how each feature contributes to the prediction for the particular instance. In techniques like Local Interpretable Model-agnostic Explanations (LIME), the probability estimate from the underlying model is used as a measurement of certainty; consequently, the feature weights represent how each feature contributes to the probability estimate. It is, however, well-known that probability estimates from classifiers are often poorly calibrated, i.e., the probability estimates do not correspond to the actual probabilities of being correct. With this in mind, explanations from techniques like LIME risk becoming misleading since the feature weights will only describe how each feature contributes to the possibly inaccurate probability estimate. This paper investigates the impact of calibrating predictive models before applying LIME. The study includes 25 benchmark data sets, using Random forest and Extreme Gradient Boosting (xGBoost) as learners and Venn-Abers and Platt scaling as calibration methods. Results from the study show that explanations of better calibrated models are themselves better calibrated, with ECE and log loss for the explanations after calibration becoming more conformed to the model ECE and log loss. The conclusion is that calibration makes the models and the explanations better by accurately representing reality.

While local explanations for AI models can offer insights into individual predictions, such as feature importance, they are plagued by issues like instability. The unreliability of feature weights, often skewed due to poorly calibrated ML models, deepens these challenges. Moreover, the critical aspect of feature importance uncertainty remains mostly unaddressed in Explainable AI (XAI). The novel feature importance explanation method presented in this paper, called Calibrated Explanations (CE), is designed to tackle these issues head-on. Built on the foundation of Venn-Abers, CE not only calibrates the underlying model but also delivers reliable feature importance explanations with an exact definition of the feature weights. CE goes beyond conventional solutions by addressing output uncertainty. It accomplishes this by providing uncertainty quantification for both feature weights and the model’s probability estimates. Additionally, CE is model-agnostic, featuring easily comprehensible conditional rules and the ability to generate counterfactual explanations with embedded uncertainty quantification. Results from an evaluation with 25 benchmark datasets underscore the efficacy of CE, making it stand as a fast, reliable, stable, and robust solution.