Towards Label Embedding -- Measuring classification difficulty

Uncertainty quantification in machine learning is a timely and vast field of research. In supervised learning, uncertainty can already occur in the very first stage of the training process, the labelling step. In particular, this is the case when not every instance can be unambiguously classified. The problem occurs for classifying instances, where classes may overlap or instances can not be clearly categorised. In other words, there is inevitable ambiguity in the annotation step and not necessarily a 'ground truth'. We look exemplary at the classification of satellite images. Each image is annotated independently by multiple labellers and classified into local climate zones (LCZs). For each instance we have multiple votes, leading to a distribution of labels rather than a single value. The main idea of this work is that we do not assume a ground truth label but embed the votes into a K-dimensional space, with K as the number of possible categories. The embedding is derived from the voting distribution in a Bayesian setup, modelled via a Dirichlet-Multinomial model. We estimate the model and posteriors using a stochastic Expectation Maximisation algorithm with Markov Chain Monte Carlo steps. While we focus on the particular example of LCZ classification, the methods developed in this paper readily extend to other situations where multiple annotators independently label texts or images. We also apply our approach to two other benchmark datasets for image classification to demonstrate this. Besides the embeddings themselves, we can investigate the resulting correlation matrices, which can be seen as generalised confusion matrices and reflect the semantic similarities of the original classes very well for all three exemplary datasets. The insights gained are valuable and can serve as general label embedding if a single ground truth per observation cannot be guaranteed.