Reliable classification with simplices

What is it about?

In a classification task, class labels for unknown data points have to be predicted based on known examples. However, for many applications, not only the predicted labels but also their predicted probability is of interest. This probability can be understood as a measure of confidence: the probability of each class label indicates the certainty of the label being the correct prediction. Classification models that provide reliable probability estimates are called well-calibrated. We propose a novel classification approach that is particularly designed to provide well-calibrated models. Our approach is based on a geometric latent space that is induced by a regular simplex with a dimension that depends on the number of classes. The visualization of the latent spaces allows us to get additional insights into the model behavior and structure of the training data.

Featured Image

Photo by Ash Edmonds on Unsplash

Photo by Ash Edmonds on Unsplash

Why is it important?

In many real-world scenarios, well-calibrated classifiers are crucial. For example, consider that in a clinical application, a machine learning model is used to predict a diagnosis from patient data. Then it becomes clear that for reliable use of such a model, it is important to know how certain a diagnosis prediction is. In particular, there may be cases in which different diagnoses apply with a similar probability.

Perspectives