Font Size: a A A

The Improvement And Research Of Locally Linear Embedding Algorithm

Posted on:2017-03-28Degree:MasterType:Thesis
Country:ChinaCandidate:M H WuFull Text:PDF
GTID:2348330488970818Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
Since the turn of the century, manifold learning as a kind of effective data dimension reduction method is more and more get the attention of scholars. Locally linear embedding (LLE) algorithm is a very effective way of data dimension reduction. It can effectively solve the problem of data "dimension disaster". But it will not be able to deal with sparse, heterogeneous data sets. In order to improve the effect of dimension reduction, we put forward the improvement method based on local linear embedding algorithm.The main research points are as follows:First of all, this paper puts forward the local neighborhood optimization method. It's based on gaussian kernel function.And the neighborhood selection on the traditional method is too big or too small. So we introduce the gaussian kernel function to you. It can combines with sample category information in the kernel space. In this way, the local neighborhood selection method is improved.Secondly, we put forward the improvement of weight vector because refactoring weight vector selection influence the result of the algorithm. But traditional locally linear embedding algorithm in computing refactoring weight vector, without considering the sample density and manifold curvature. So we plan the selection of weight vector for further improvement. By this way, the algorithm has strong robustness.Finally, our algorithm is further improved. In the traditional local linear embedding algorithm, the high dimensional spatial data is embedded into the low dimensional space. So the weight vector is only determined by the source of high dimensional data space. However, considering the weight vector can be determined by the high-dimensional source and low dimensional data space together. And the experimental results show the effectiveness of the improved method.
Keywords/Search Tags:manifold learning, locally linear embedding, dimension reduction, local neighborhood
PDF Full Text Request
Related items