Sparse nonlinear dimension reduction algorithm

Abstract

After we go into the big data age, the scale of data sets becomes more and more massive. The dimension of the data is also higher. For a task, there are a huge number of features in the ultra-high-dimensional data set that are irrelevant to the learning task. Manifold learning is proposed to reduce the dimensions of high-dimensional data sets and to use a manifold and other geometric tools to extract features from the data sets to obtain valid features and eliminate invalid features. However, the traditional manifold learning algorithms still have some unresolved problems by themselves. At present, the classic manifold learning algorithm can not deal with some large scale data sets. The main reason is that the cost of memory is very high during the algorithm construction. Therefore, this article makes some improvement for the basic manifold algorithm and proposes a sparse algorithm to reduce the memory cost and the time complexity. In experiments, we are using three sets of combinational data sets for comparative experiments. The experimental results show that the improved algorithm has obvious advantages in terms of time consumption.