Comparative Research On Methods Of Dimensionality Reduction In Medical Data

With the development of information and network technology, there appeareda large number of high dimensional data in every walks of life, for example:medical data, text data, network data and financial data and so on. Highdimensional data brought us “Dimension Gospel”, at the same time it also broughtus “Dimension Disaster”. So how to express the high dimensional data in lowdimensional space and find out the inner structure had become the key to deal withthe high dimensional data. Dimension reduction, as an effect method to convert“Dimension Disaster” into “Dimension Gospel”, has caused people’s attentionwidely, and been developing. The effect of linear dimension reduction method isvery good in dealing with linear data sets in global. There is high requirement innonlinear data sets for nonlinear dimension reduction method, such as sensitivity tonoise and the dense of samples, so the effect of nonlinear dimension reductionmethod is good in ideal data sets. Actually it is not always better than traditionallinear dimension reduction method in actual data sets.Firstly, this paper introduces the background of dimension-reduction methods,the research significance and the research status. Secondly， this paper introducesstatistic linear dimension-reduction method(PCA) and nonlineardimension-reduction method(Isomap and LLE). At the same time it proposed thebig correlation PCA algorithm(BR-PCA) after improving the way to select principalcomponents in the base of PCA algorithm. And we analyzed and compared the fourmethods from some aspects, for example: algorithm thought, time complexity ofalgorithm, requirement for data sets, target dimension, theoretical support and theexplanation of the result of dimension-reduction. We reduced the high dimensionalcancer data sets in four methods(target dimension is equal), then predicted andcompared the prognosis diagnosis of five-year survival of colorectal cancer patientwith the same BP Neural Network. We found that the linear dimension-reductionmethod(BR-PCA algorithm) is better than nonlinear dimension-reduction method indealing with real nonlinear data sets after the comparing of the result, overcamesome weakness of Manifold Learning, at the same time it is better than traditional linear dimension-reduction method(PCA algorithm) in prediction and analysis, too.At last, the conclusion was tested in different preinstalled error and human facedata sets.