A Prescriptive Machine Learning Approach for Assessing Goodwill in the Automotive Domain

Abstract

Car manufacturers receive thousands of goodwill requests for vehicle defects per year. At BMW, these requests for repair-cost contributions are either assessed automatically by a set of fixed rules or manually by human experts. To decrease manual effort, which is still around 50%, we propose a machine learning approach with the goal to discover so far unknown assessment patterns in human decisions. Since the assessment contribution data is heavily imbalanced, we structure the learning task hierarchically: The first layer’s task is to predict the main rank of the request (no contribution, partial contribution, or full contribution). Then, in the case where partial contribution is suggested, the second layer predicts the concrete percentage using a regression model. To optimize our model and tailor it to certain strategies (e.g., customer friendly or more cost oriented), we make use of a custom-defined cost matrix. We also outline how the model can be used in a scenario in which it prescribes appropriate monetary contributions for requested repair-costs. This can initially happen in the form of a decision support system (DSS) and, in the next step, through automated decision making (ADM), where a certain part of goodwill requests is processed automatically by the prescriptive model.