PROJECT TITLE :

Progressive Self-Supervised Clustering With Novel Category Discovery

ABSTRACT:

Clustering is one of the most classical methods used to analyze data structures in machine learning and pattern recognition at the moment. In recent years, the anchor-based graph has gained widespread adoption as a means of enhancing the accuracy of clustering performed by numerous graph-based clustering techniques. The progressive self-supervised clustering method with novel category discovery (PSSCNCD) is a novel clustering approach that we propose in order to achieve more satisfying clustering performance. This method, which specifically consists of three different procedures, is known as the progressive self-supervised clustering method. First, we present a brand new semisupervised framework that makes use of novel category discovery to direct label propagation processing. This framework is supported by the parameter-insensitive anchor-based graph that is obtained from balanced K-means and hierarchical K-means, respectively (BKHK). Second, we design a novel representative point selected strategy based on our semisupervised framework to discover each representative point and endow pseudolabel progressively. Each pseudolabel hypothetically corresponds to a real category in each self-supervised label propagation. This allows us to discover each representative point and endow pseudolabel progressively. Third, once a sufficient number of representative points have been located, the labels of all of the samples will be finally predicted in order to obtain the terminal results of the clustering. In addition, the results of the experiments performed on a number of dummy examples as well as benchmark data sets provide conclusive evidence that our method is superior to other clustering approaches.