فهرست نشریات علمی دسترسی آزاد ایران

With regard to the non-linear nature of real-life data, their clusters' shapes are non-convex and unfortunately, some clustering methods cannot identify non-convex clusters and this is a challenge. Density-based clustering methods could be a solution to this problem. Among all methods of this type, the DBSCAN algorithm can cluster data with different shapes, sizes, and densities and also identify noise points. However, owing to the use of static input parameters-the neighbourhood radius ($Eps$) and the minimum value for cluster formation ($MinPts$)- this algorithm has some problems such as the difficulty in accurately determining these parameters in high-dimensional data sets and not recognizing clusters with different densities. Accordingly, this paper presents a density clustering algorithm, which requires minimal input parameters and one of its main parameters is $Eps$, which is automatically calculated based on the $k$-nearest neighbours of points and its value is different for each cluster. To evaluate the effectiveness of the proposed algorithm, some experiments were conducted. The obtained results showed the effectiveness and efficiency of the presented algorithm regarding the correct identification of clusters with the desired shape, size, and density. In addition, the proposed algorithm was found effective in estimating the number of clusters in most of the data sets considered in this study.