Abstract:A de-noising hierarchical mapping (DHM) algorithm is proposed in order to use effective deep feature selection methods in multi-dimensional clustering analysis to eliminate redundant and irrelevant features and learn the nonlinear relationship of data elements to extract the best features. In the algorithm, an acyclic neural network is first built based on the denoising autoencoder. Specifically, the fault-tolerant data are trained by hidden layer weighting and activation function to obtain the nonlinear relationship of the input data and the feature space. The features are reconstructed and the best features are selected. Secondly, the feature space is used to adjust the self-organizing feature map neural network. the minimized weighted squared Euclidean distance is calculated to find the matching winning neuron. Finally, the feature selection network and the unsupervised clustering network are combined to construct the noise reduction hierarchical map neural network. The noise reduction hierarchical map neural network is iteratively trained and the weight parameter and the deviation vector are optimized at the same time to realize an effective unsupervised clustering scheme. Experimental results on real data sets show that, compared with AESOM, DCSOM and S-SOM algorithms, the DHM algorithm has better performance in the quality and accuracy of clustering.
ICP: