Estimating Multi-label Accuracy using Labelset Distributions

A multi-label classifier estimates the binary label state (relevant vs irrelevant) for each of a set of concept labels, for any given instance. Probabilistic multi-label classifiers provide a predictive posterior distribution over all possible labelset combinations of such label states (the powerset of labels) from which we can provide the best estimate, simply by selecting the labelset corresponding to the largest expected accuracy, over that distribution. For example, in maximizing exact match accuracy, we provide the mode of the distribution. But how does this relate to the confidence we may have in such an estimate? Confidence is an important element of real-world applications of multi-label classifiers (as in machine learning in general) and is an important ingredient in explainability and interpretability. However, it is not obvious how to provide confidence in the multi-label context and relating to a particular accuracy metric, and nor is it clear how to provide a confidence which correlates well with the expected accuracy, which would be most valuable in real-world decision making. In this article we estimate the expected accuracy as a surrogate for confidence, for a given accuracy metric. We hypothesise that the expected accuracy can be estimated from the multi-label predictive distribution. We examine seven candidate functions for their ability to estimate expected accuracy from the predictive distribution. We found three of these to correlate to expected accuracy and are robust. Further, we determined that each candidate function can be used separately to estimate Hamming similarity, but a combination of the candidates was best for expected Jaccard index and exact match.

Radio Galaxy Zoo: Unsupervised Clustering of Convolutionally Auto-encoded Radio-astronomical Images

This paper demonstrates a novel and efficient unsupervised clustering method with the combination of a Self-Organising Map (SOM) and a convolutional autoencoder. The rapidly increasing volume of radio-astronomical data has increased demand for machine learning methods as solutions to classification and outlier detection. Major astronomical discoveries are unplanned and found in the unexpected, making unsupervised machine learning highly desirable by operating without assumptions and labelled training data. Our approach shows SOM training time is drastically reduced and high-level features can be clustered by training on auto-encoded feature vectors instead of raw images. Our results demonstrate this method is capable of accurately separating outliers on a SOM with neighbourhood similarity and K-means clustering of radio-astronomical features complexity. We present this method as a powerful new approach to data exploration by providing a detailed understanding of the morphology and relationships of Radio Galaxy Zoo (RGZ) dataset image features which can be applied to new radio survey data.