Research on Improved Adaptive Spectral Clustering HRM-ISC Algorithm

As a crucial component in the realm of data mining, In the past few years, the focus of researchers in the field of data mining has significantly shifted towards the exploration of clustering analysis, garnering widespread interest and recognition, and a considerable multitude of theories and methods have been produced. Based on the conventional clustering algorithm, spectral clustering converts the partition of the data set into the partition of the graph, and can identify the clusters composed of arbitrary shape sample data. The traditional spectral clustering algorithm NJW has problems such as difficult parameter selection in the clustering process and poor clustering effect on non-convex or non-uniform density data sets. In this paper, a hot region mining algorithm built upon improved spectral clustering (HRM-ISC) is suggested. In the first place, The enhancement of the local scale parameter in the self-tuning algorithm is achieved by incorporating the standard deviation of the sample neighborhood, which avoids the interference of the local scale parameter of the sample by noise points and the parameter selection depends on the empirical value to the maximum extent, so that the similarity matrix of the sample reflects the distribution information of the sample more accurately. Secondly, The conventional spectral clustering approach can not automatically estimate the optimal number of clusters and the eigenvector selection is difficult. According to the matrix perturbation theory, an adaptive spectral clustering algorithm utilizing eigengap is used to solve the problem. Finally, to tackle challenge of sensitivity to the initial clustering center encountered in the K-Means algorithm, the K-Medoids algorithm with variance maximization of initial clustering center was employed as a substitute.