Improving baseball pitch type classification through clustering of pitchers

Abstract

In this paper, we propose a method for improving the performance of baseball pitch type classification through clustering of pitchers. Previous studies assume an individual classification model for each pitcher. This approach can reflect personal characteristics

however, it can suffer from a lack of data especially for new players. This problem can be resolved using the single unified model classification

however, it is not flexible enough to capture individual differences. To retain the advantages of both approaches while complementing their disadvantages, we propose to cluster pitchers by several homogeneous groups that share similar characteristics, such as left/right handedness and pitching form, and then build separate classification models for different clusters. We show via real-data experiments that using the proposed approach, we can increase the classification accuracy with less standard error. In particular, we achieved the best performance when the multi-layer perceptron was used as a classification algorithm, in combination with k-means clustering. We also discuss the benefits of our clustering approach in characterizing pitchers for making strategic decisions for effective team management.