Unsupervised Dimensionality Reduction By Merging Model Space

Dimensionality reduction(DR) is a key issue in the field of pattern recognition, machine learning and data mining. Dimensionality reduction techniques map highdimensional data to low-dimensional space by linear or non-linear transformation so as to facilitate subsequent data analysis. However, with the rapid development of the data acquisition technology, communication technology and the computer technology, the volume of the data is bigger and bigger, the dimension is higher and higher, and this makes the computational complexity of traditional dimensionality reduction techniques very high,and even they cannot be applied. In order to achieve the dimensionality reduction of the large-scale and high-dimensional data in a fast and effective way, this article will give an in-depth study of incremental unsupervised dimensionality reduction techniques based on building and merging representation models of data. The main contents of this article are as follows:1. The existing incremental 2DPCA algorithms can not update the projection directions using multiple images, also they ignored the average of the training data. In order to solve this problem, we proposed model merging-based incremental two-dimensional principal component analysis(M-I2DPCA) algorithm which can describe the data space by building several feature spaces and then merge them to get the eigenspace model and the projection directions of the whole data. Compared to 2DPCA algorithm, M-I2 DPCA can significantly reduce the computational complexity and the storage requirement. Experimental results on the FERET, AR, and PIE face databases show that the proposed approach has the performance as good as one obtained by batch 2DPCA.2. When dealing with large-scale and high-dimensional data, LPP leads to a high computational complexity, also the storage and the decomposition of big matrices may lead to a significant performance degradation. In order to solve this problem, we proposed model merging-based incremental locality preserving projection(M-ILPP) algorithm. By building and merging different LPP representation sub-models of separated subsets of the data, M-ILPP can get the final LPP representation model and then calculate the projection directions. We demonstrate the feasibility and effectiveness of our method with experimental results on several databases.