2165. Counterfactual Explanations in Regression
Invited abstract in session WC-12: Explainability and Interpretability in Optimization, stream Artificial Intelligence, Machine Learning and Optimization.
Wednesday, 13:30-15:00Room: H10
Authors (first author is the speaker)
| 1. | Matthias Soppert
|
| Chair of Business Analytics & Management Science, University of the Bundeswehr Munich | |
| 2. | Alexandre Forel
|
| Department of Mathematics and Industrial Engineering, Polytechnique Montreal | |
| 3. | Thibaut Vidal
|
| Mathematics and Industrial Engineering, Polytechnique Montréal | |
| 4. | Maxime Cohen
|
| McGill University |
Abstract
Accurate predictions of real-valued quantities form the basis for data-driven decision-making in many applications across all industries. For example, online retailers as well as classical brick-and-mortar stores require accurate demand predictions for effective operations management, such as pricing and inventory control. Often, however, the best performing models regarding prediction accuracy are opaque and, thus, are considered as black boxes by the user. We propose a framework based on mixed-integer programming that increases the explainability of a wide range of state-of-the-art regressors, including random forests, gradient boosting regressors, and certain artificial neural networks. The framework is based on the determination of counterfactual explanations – an established approach to improve explainability. It yields answers to questions of the form "what is the minimum required change of a certain model input such that the model's prediction exceeds a specific target output?". Thus, the answers are optimal with regard to the required model input change. For example, for a given vector of product features and context information as well as its predicted sales volume, the optimization returns the closest alternative vector that results in a specific desired minimum sales volume. Asking such counterfactual questions and analyzing their respective answers helps to explain the input-output relations that a trained regressor has learned, but they can also directly be used in decision-making – in the above example possibly regarding marketing activities.
Keywords
- Machine Learning
- Mixed-Integer Programming
Status: accepted
Back to the list of papers