A semi-supervised fuzzy clustering approach via modifications of the dbscan algorithm

Abstract

In the data mining context, semi-supervised learning is applicable in circumstances where only a scarce amount of information on the intrinsic structure of a dataset is available. This information may be in the form a few labelled instances or a relatively small set of constraints on the pairwise memberships of particular instances. In this study we are providing a semi-supervised fuzzy clustering model which modifies versions of conventional DBSCAN algorithm in order to generate soft clusters which foreclose the noise points. The employed modifications are mostly related to the control parameters of the algorithm intending to utilize the additional information (which in our case is in the form of a few labelled instances) and adaptations towards the fuzzy clustering approach. Finally, several experimental procedures have been conducted on synthetic and real-world benchmark datasets in order to assess the accuracy of our employed model and to compare it to the conventional algorithms of the respective domain.