Support Vector Machines Classifier Visualization Research Based on Parametric Embedding Algorithm
Xiaofang Yuan · Journal of Hunan University · 2008
In order to solve the problem that only one or several aspects of high-dimensional data characteristics are reflected when using general dimensionality reduction methods,Parametric Embedding(PE) algorithms were applied to embedding all the characteristics in a low-dimensional visualization space.Firstly,Posterior probability of training data over classes in high-dimensional space was obtained by Support Vector Machines(SVM) classification algorithms based on the assumption that the class of data can be represented by a Gaussian distribution in the embedding space.Then,PE was applied to obtain the test data coordinates and class properties in the embedding space,thus implementing high-dimensional data classifier visualization.Experimental results on USPS,YaleB,MiniNewsgroups data sets show the method can represent not only the relations between data and classes,but also the relations within the set of data and within the set of classes,which is beneficial to deeply study on high-dimensional data classifying properties.The visualization effects of PE are better than those by directly applying PCA,ISOMAP algorithms.In addition,the PE computing complexity is just the product of the number of objects and the number of classes,which is suitable to large volume of data,fewer types of classified data visualization.