A Deep Non-Parametric Learning Architecture for High Dimensional Data Clustering- Deep Learning Paradigm

Abstract

Data Clustering is a fundamental research focus in many data-driven high dimensionality application domains and clustering performance of data with high dimensionality is highly depends on the quality of data representation in terms of effectiveness and efficiency. Machine learning is a classical approach to provide soft partition of data but faces many difficulties due to increasing sparsity of data and increasing difficulties in distinguishing distance between data points, especially it is high complex in handling high dimensional data with complex latent distributions. In order to tackle those issues, a new deep learning paradigm has been presented in this paper. The proposed model is termed as deep non parametric learning architecture for high dimensional data clustering. The Proposed model leverages unsupervised feature learning along autoencoder for high cluster friendly representation. It first map the data into embedding objects and embedded objects has been  taken for clustering on generation of objective function to produce maximum margin cluster. Those cluster further fine tuned to refine parameter of different layers of convolution neural network to ensure the minimum reconstruction error between feature max pooling layer and ReLu activation layer. Softmax layer minimizes the intra cluster compactness and inter cluster seperability in the feature space. Non parametric tuning has been enabled in the output layer to make the data instance in the cluster to be close to each other by determining the affinity of the data on new representation. It results significant increase in the clustering performance on the discriminative information’s. Extensive experiments have been conducted on real datasets to compare proposed model with several state-of-the-art approaches. The experimental results show that Deep Non-Parametric learning architecture can achieve both effectiveness and good scalability on high dimensional data.