Interactive visual labelling versus active learning: an experimental comparison

Methods from supervised machine learning allow the classification of new data automatically and are tremendously helpful for data analysis. The quality of supervised maching learning depends not only on the type of algorithm used, but also on the quality of the labelled dataset used to train the classifier. Labelling instances in a training dataset is often done manually relying on selections and annotations by expert analysts, and is often a tedious and time-consuming process. Active learning algorithms can automatically determine a subset of data instances for which labels would provide useful input to the learning process. Interactive visual labelling techniques are a promising alternative, providing effective visual overviews from which an analyst can simultaneously explore data records and select items to a label. By putting the analyst in the loop, higher accuracy can be achieved in the resulting classifier. While initial results of interactive visual labelling techniques are promising in the sense that user labelling can improve supervised learning, many aspects of these techniques are still largely unexplored. This paper presents a study conducted using the mVis tool to compare three interactive visualisations, similarity map, scatterplot matrix (SPLOM), and parallel coordinates, with each other and with active learning for the purpose of labelling a multivariate dataset. The results show that all three interactive visual labelling techniques surpass active learning algorithms in terms of classifier accuracy, and that users subjectively prefer the similarity map over SPLOM and parallel coordinates for labelling. Users also employ different labelling strategies depending on the visualisation used.