Linking Data Separation, Visual Separation, Classifier Performance Using Multidimensional Projections

Understanding how data separation (DS), visual separation (VS), and classifier performance (CP) are related to each other is important for applications in both machine learning and information visualization. A recent study showed that, for a specific machine learning pipeline using a given multidimensional projection technique, high DS leads to high VS and next high CP. However, whether such correlations would stay the same (or not) when using other projection techniques was left open. We fill this gap by evaluating ten projection techniques in a pipeline that uses three contrastive learning methods (SimCLR, SupCon, and their combination) to produce latent spaces and next train and test classifiers for five image datasets of real-world application with human intestinal parasites. Our work identifies two classes of projection techniques – one leading to poor VS and next poor CS regardless of the available DS, and the other showing a good DS-VS-CP correlation. We argue that this last group of projections is a useful instrument in classifier engineering tasks.