In recent years, supervised machine learning models trained on videos of animals with pose estimation data and behavior labels have been used for automated behavior classification. Applications include, for example, automated detection of neurological diseases in animal models. However, there are two problems with these supervised learning models. First, such models require a large amount of labeled data but the labeling of behaviors frame by frame is a laborious manual process that is not easily scalable. Second, such methods rely on handcrafted features obtained from pose estimation data that are usually designed empirically. In this paper, we propose to overcome these two problems using contrastive learning for self-supervised feature engineering on pose estimation data. Our approach allows the use of unlabeled videos to learn feature representations and reduce the need for handcrafting of higher-level features from pose positions. We show that this approach to feature representation can achieve better classification performance compared to handcrafted features alone, and that the performance improvement is due to contrastive learning on unlabeled data rather than the neural network architecture.