关键词: DBSCAN算法, 去噪衰减, 多密度阈值, 自适应

Abstract: The parameters of [Eps] and [MinPts] in DBSCAN algorithm need to be set artificially. Improper values will result in low accuracy of clustering results. Moreover, due to the global nature of the parameters, they are mistakenly applied to clusters with different densities, causing the cluster not to be found correctly. To solve above problems, this paper puts forward a kind of multi-density adaptive parameter determination algorithm, the algorithm first generates candidate [Eps] and [MinPts] parameter lists by using the self-distribution characteristics of the denoising attenuated data set, and in the interval where the number of clusters tends to be stable, corresponding [Eps] and [MinPts] are selected as the initial density threshold according to the denoising level. Then, the candidate parameter list is generated by using the same method for noise data generated by clustering under the condition of the density threshold, and the optimal parameter is selected to obtain the new density threshold. The step is cycled until the noise data quantity or density threshold is lower than a certain degree. Finally, the clustering results of different density thresholds are combined to obtain the final clustering result. The experimental results show that the algorithm can adaptively select the appropriate multi-density threshold and has a good clustering effect on the data sets with large density distribution differences.

Key words: DBSCAN algorithm, denoising attenuation, multi-density threshold, self-adaptive