In manufacturing, multi-criteria decision making (MCDM) methods are already widely known and increasingly applied to multi-objective optimization problems. Oftentimes though, as in machine scheduling, conditions surrounding these problems remain constant, such as the number of available machines or the responsible decision maker (DM) assigning tasks to these machines. Therefore, for frequently changing job demands MCDM method users repeatedly solve very similar scheduling problems, while each instance of the MCDM method can require a large number of user interactions of the DM. Accordingly, vast historical user interaction and preference data is generated. One interactive MCDM method that enables DMs to find near optimal solutions for three-objective machine scheduling problems is the Interactive Centroid Method (ICM). In this study, an attempt is made to create a preference learning method to complement the ICM. For a single user, the underlying unknown preference function remains constant and can be learned from historical user interaction data. The proposed SVM-based preference learning method can remain within a reasonable computational time, as well as retain the low number of user interactions that the ICM requires. Thereby, future demand scenarios can be solved preemptively and a solution is suggested to the decision maker. This approach has the potential to drastically reduce user decision time by replacing an average of 10 user interactions for each problem instance with one single final validation of the suggested solution. The proposed SVC (Support Vector Classification) method is tested for linear and quadratic preference functions, as well as varying amounts of available historical preference data. Especially for quadratic preferences, the SVC results are in most cases an adequate replacement for further applications of the ICM.