Research On Incremental Manifold Learning

Dimensionality reduction is a key issue in the fields of pattern recognition, machine learning and data mining. It aims to map high-dimensional samples into low-dimensional subspace by linear or nonlinear transformation, which facilitates the subsequent analysis. With the rapid development of the data acquisition technology, communication technology and the computer technology, the amount of the sample is bigger and the dimension is higher, which makes the computational complexity of traditional dimensionality reduction techniques become very high, even cannot be applied. In order to alleviate this problem, this article starts with manifold learning, and then gives an in-depth study of incremental unsupervised dimensionality reduction techniques based on manifold learning methods. The main contents of this article are as follows:1. When addressing a new sample in the batch mode, Neighborhood preserving embedding algorithm has a high computational complexity. In order to alleviate this problem, this article gives a research on the incremental neighborhood preserving embedding(INPE) algorithm. When a new sample is added, the incremental learning algorithm uses the previous processed results and avoids the repeated operation. In addition, INPE reduces the computational complexity, also improves the processing speed and realizes the incremental processing of NPE algorithm. The experimental results on the AR, ORL and FERET face databases validate the feasibility and effectiveness of this method.2. When dealing with large-scale and high-dimensional samples, 2DLPP has a high time complexity and space complexity. In order to alleviate this problem, this article gives a research on the model merging-based incremental two-dimensional locality preserving projection(M-I2DLPP) algorithm. By building and merging different 2DLPP representation sub-models of separated subsets of the samples, M-I2 DLPP can get the final 2DLPP representation model and then calculate the projection directions by the decomposition of matrix. It demonstrates the feasibility and effectiveness of this method with experimental results on AR, ORL and FERET face databases.