E-DBSCAN: An evidential version of the DBSCAN method

In later years, data have grown enormously and dealing with them to extract information has become a necessity. Data mining is a subfield of both computer science and statistics that aim to extract useful information in a comprehensive structure. The importance of clustering techniques in data mining has lead to the development of many methods in order to deal with data. Among these methods, we name density-based techniques, such as DBSCAN, that partitions data into heterogeneous shapes according to their local densities. DBSCAN can be suitable when handling big data that have noises and outliers. However, the classic DBSCAN method fails in identifying clusters with a variable density distribution and overlapping borders which is accurate in real-world data. In this paper, we propose an unsupervised learning technique in an uncertain context, that combines the DBSCAN method and the framework of the belief function theory, in order to generate clusters having overlapping borders. The proposed evidential clustering method, that we called E-DBSCAN, has the ability to handle cluster membership degree uncertainty of objects by using the belief function theory.