A GA Based Approach to Improving the ICA Based Classification Models for Tumor Classification
A GA Based Approach to Improving the ICA Based Classification
Models for Tumor Classification
Jian-Bing Xia-Hou School of Software Xiamen University Xiamen, Fujian Province
China
jbxiahou@xmu.edu.cn
Kun-Hong Liu* (Corresponding Author)
School of Software Xiamen University Xiamen, Fujian Province
China
lkhqz@xmu.edu.cn
Abstract: - As it has been pointed out that different ICs are of different biological significance, this paper tries to explore the IC selection problem based on a set of experiments. A regression model and a classification model, referred as penalized independent component regression (P-ICR) and ICA based Support Vector Machine (ICA+SVM), are applied to illustrate the necessity and efficiency of IC selection. A genetic algorithm (GA) is deployed to the selection process, along with an early stopping technique deployed to avoid overfitting in evolution. In particular, the individuals in the selected generation are used to construct an ensemble system to achieve higher classification accuracy. We test the two models with and without the selection methods based on three microarray datasets. The experiment results demonstrate that IC selection methods can further improve the classification accuracy of the ICA based prediction models, and the GA is more effective than the original methods.
Key-Words: - Microarray data; Independent component analysis (ICA); Genetic algorithm (GA); Early stopping; Support Vector Machine(SVM); Overfitting;
has the great potential advantages over PCA in 1 Introduction
many aspects [2]. As a result, ICA technique has With the development of microarray technology,
been attracting more and more attentions so far. it is possible to diagnose and classify some
To illustrate a typical ICA processing, assume particular cancers directly based on DNA
that an n p data matrix X, with rows ri (i=1,…,n) microarray data. However, the class prediction of
corresponding to observational variables and microarray data is a typical “large p, small n”
columns cj ( j=1,…,p), is the individual of the problem [1], which means that the number of
corresponding variables, the ICA model of X can be predictor variables is much larger than the number
written as: of samples. So it is a great challenge to develop a
X=AS (1) new efficient method for analyzing global gene
Without loss of generality, A is an n n matrix, expression data. Currently a variety of algorithms
and S is an n p source matrix. Those variables in and mathematical models have been used for
the rows of S are ICs, and the statistical management, analysis and interpretation of these
independence between variables can be quantified high-density microarray data.
by mutual information I=∑H(Sk)-H(S), where H(Sk) Principal component analysis (PCA) is a widely
is the marginal entropy of the variable Sk, and H(S) deployed tool for finding useful eigenassay or
is the joint entropy. And estimating the independent eigengene. With its aid, the gene snapshot
components can be accomplished according to the coordinates can be predicted from each other
formula: linearly. But independent component analysis (ICA)
U=S=A-1X=WX (2)
A Method of Soil Salinization Information Extraction with SVM Classification Based on ICA and Te_专业资料。Arutrlcne&Tcnly,0127:06—09,04gilaSiccu eehoog...
A Method of Soil Salinization Information Extraction with SVM Classification Based on ICA and Te_专业资料。Arutrlcne&Tcnly,0127:06—09,04gilaSiccu eehoog...
热门文档