Interactive clustering: a scoping review

Abstract

We present in this paper a scoping review conducted in the interactive clustering area. Interactive clustering has been applied to leverage the strengths of both unsupervised and supervised learning. In interactive clustering, supervised learning is represented by inserting the knowledge of human experts in an originally unsupervised data analysis process. This scoping review aimed to organize the knowledge on (i) the applicability of interactive clustering methods, (ii) clustering algorithms being used to support interactive clustering, (iii) how to model the expert supervision and (iv) the effects brought by the expert supervision in the results produced. A systematic search for related literature was conducted in the Scopus database, resulting in the selection of 50 primary studies published by 2018. The analysis of these studies allowed us to identify trends such as: the application in text/image; use of partitioning and hierarchical algorithms; application of strategies based on split/merge, pairwise constraints, similarity metrics learning and data reassignment; and concern with visualization. In addition, some relevant issues not yet adequately addressed were identified, such as: the evaluation of expert supervision; the evaluation of the expert’s effort; and the conduction of studies effectively involving human experts, instead of computer simulations.