HAD-UMAP: Hybrid Attribute Data Set with Uniform Manifold Approximation and Projection for Adaptive Spectral Clustering

In the clustering task, there are many mature algorithms. Due to the high dimension of the current data, if the appropriate dimensionality reduction is not carried out, the noise of the data will easily affect the accuracy of the algorithm. Meanwhile, to solve the attribute skew problem of spectral clustering algorithm when processing hybrid attribute data set, and the artificial selection problem of scale parameters in Gaussian kernel function, this paper proposes a novel hybrid attribute data set with uniform manifold approximation and projection (UMAP) for adaptive spectral clustering. UMAP method is used to reduce the dimensionality of the original data, and regularization spectral clustering algorithm is used to decompose the spectrum. This new method improves the traditional classification attribute similarity measure by calculating the information entropy of numerical attribute and classification attribute and obtaining the balance difference factor. In Gaussian kernel function, the neighborhood radius of each sample is calculated by using shared natural neighbors, and the scale parameters are solved adaptively. Finally, the similarity matrix of hybrid attribute samples is constructed by kernel function for spectral clustering. The experimental results show that proposed algorithm is better than four common hybrid attribute data clustering algorithms in ACC, ARI and NMI indexes. The clustering results are more stable. The new algorithm can effectively solve the attribute skew problem, especially the real distribution information of the hybrid attribute data set can be found completely adaptively, and the clustering efficiency is significantly improved.