Cluster-clean-label
David Beil, Andreas Theissler · 2020
One of the major problems of applying supervised machine learning methods in real-world problems is the absence of labeled data. Labeling huge amounts of data is time consuming and cost intensive. Moreover, in many cases, labels can only be assigned by domain experts like medical doctors or engineers, who have little time and do not necessarily have profound machine learning knowledge. In this paper, we propose an efficient interactive cluster-clean-label approach. First, to visualize the potentially huge amount of data, principal component analysis followed by t-SNE projection is applied. On the 2-dimensional representation of the data, HDBSCAN clustering is utilized to identify groups of potentially similar class membership. Subsequently, anomaly detection in form of an autoencoder is applied on each cluster, and instances that are likely to belong to different classes are suggested to the user. The user decides which of these suggested instances to include and restarts the anomaly detection process with the remaining subset of instances. This iterative process is repeated until the user is satisfied with the clusters' purity. Eventually, labels are assigned to the clusters. The approach is evaluated by a user study with 25 participants using the initially unlabeled MNIST data set, where on average users were able to label 91.59% of the data set, with an accuracy of 98.99%. A video showing the approach is available: https://youtu.be/RsLI0dg90qE.