High Dimensional Deep Data Clustering Architecture Towards Evolving Concept

Main Article Content

A. Sumathi , K. Yasotha , S. Nandhinidevi

Abstract

 Data Clustering on the evolving data stream has becoming primary and vital tasks in the natural language processing in data driven application domains. Performance analysis of data clustering techniques depends on the quality of data representation on the evolving data streams. Machine learning algorithm plays a primary role in high dimensional data clustering of the evolving data streams from social media. Despite of large benefits on implementing those algorithms, it suffers on various aspects such as low accuracy and distribution of the data points in the time varying clusters. In order to resolve above mentioned issues, Deep learning architectures has been analysed on various aspects to develop a optimized model for better data representation of data clustering. In this paper, an optimized deep learning architecture which termed as Concept Based High Dimensional Deep Clustering has been projected for analysing unstructured and high dimensional evolving data in the data streams. Proposed architectures use the hidden layer for better effective identification of the hidden feature. Further the evolving data uses transform learning for data transformation. Those learning transforms high dimensional to low dimensional deep feature space. Finally these deep feature spaces extract the concept specific features on the employment of the max layer using principle component analysis (PCA). PCA on the max layer determines the salient features on ensuring the minimum reconstruction error. Deep architecture is eliminating the NP hard problem and over fitting issues of the clustered results. Further all parameters are fine tuned with respect to certain criterion on cross validation. The Softmax layer is used to map the data points into accurate cluster representations. Finally it is helpful to find a better initialization of the parameters. Extensive experiments have been conducted on real datasets to compare proposed model with several state-of-the-art approaches. The experimental results show that proposed deep clustering model can achieve both effectiveness and good scalability on high dimensional data.

Article Details

Section
Articles