CBR-fox: A Case-Based Explanation Method for Time Series Forecasting Models.

Explainable Artificial Intelligence refers to methods that help human experts understand solutions developed by Artificial Intelligence systems in the form of black-box models, making them transparent and understandable. This paper describes CBR-fox, a post-hoc model-agnostic case-based explanation method for forecasting models. This method generates a case base of explanation examples through a sliding-window technique applied over the time series. Then, these explanation cases can be retrieved using a wide range of well-established metrics for time series comparison. Moreover, we introduce and evaluate a novel similarity metric named Combined Correlation Index. The proposed retrieval approach considers as a signal the similarity series resulting from applying the comparison metrics. This way, the signal can be smoothed using noise removal filters, such as the Hodrick-Prescott and low-pass filters, to avoid maximally similar cases that may overlap or represent a local slice of the source time series.. The resulting signal allows then to foster diversity in the retrieved explanation cases presented to the user The proposed case-based explanation approach is evaluated in the weather forecasting domain using an artificial neural network as the black-box model to be explained.