文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > IT/计算机 > The Analysis of Microarray Datasets Using a Genetic Programming
侵权投诉

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

promising solution for the discovery of potentially important gene expression profiling data because it can generate comprehensible rules for classification from microarray datasets owing to its powerful search ability. The advantage of GP is that it can solve the feature selection process and classification task simultaneously, and also provide deep insight of the biological significance of genes.

This paper is organized as follows. In Section II, we present the application of GP algorithm in construction of classifying rules. The dynamic maximum tree depth technique and weight-based scheme are used in this procedure. In Section III, the experimental designs and results are presented. Some biological analyses are also included in this section.

II. METHODS

GP was introduced by Koza (Koza, 1992) [6]. It is a branch of genetic algorithm (GA). And the main difference between GP and GA is the representation of the solution. GA has string-structured individuals, which usually represent different feature masks for a candidate feature set when dealing with the feature selection problem; GP’s individuals are trees, in which the leaf nodes and non-leaf nodes are chosen from the terminal and function sets, respectively. Usually, in the feature selection problem, the terminal set consists of features and constants, and the function set consists of several mathematical or logical functions, which are used to connect different terminals.

Owing to the individual structure, GP can produce classification rules by formulating important features. Generally, in the feature selection problem, the terminal set consists of features and constants, and the function set consists of arithmetical or logical functions. The leaf nodes and non-leaf nodes of trees are chosen from the terminal and function sets, respectively. Let F be the set of functions, and T be the set of terminals. The trees built in the evolution process are the set of all possible compositions of functions and terminals selected from F and T. When used for the classification task, a tree is evolved from a training dataset and validated against an independent test dataset. Take the rule for distinguishing Leukemia dataset as an example, which is listed in Table 2 as: ‘if gt(X3665, X2515) then ALL’. Here, the serial number of a gene is indicated by the letter ‘X’ followed by a number. So X3665 refers to the 3665th gene in the Leukemia dataset. The ‘if’ clause is generated by the GP, that is, the function ‘greater’ and two genes are selected by GP. The target class is predefined as the ALL. According to the function settings shown in Table 1, function ‘greater’ performs comparison on the expression levels of these two genes. Only when the gene expression value of the 3665th gene is greater than that of the 2515th gene, the tree produces an answer ‘yes’ (+1) and assigns this sample to the ALL; otherwise, it produces ‘no’ (-1). All trees can be comprehended in this way, regardless of their depths.

The Ramped Half-and-Half method proposed in (Koza, 1992) is applied here to generate the first generation. In this

procedure, an equal number of trees are initialized for each depth between 2 and the initial dynamic tree depth value. For each depth level considered, half of the individuals receive non-leaf nodes from F until trees are fully grown; the other half are allowed to receive nodes from both F and T randomly except for the root node, which produces a set of heavily unbalanced trees. Then this method results in balanced and unbalanced trees with several different depths.

The dynamic maximum tree depth technique proposed in (Silva and Almeida, 2003) [7] is used to control bloat. With this technique, two different parameters are used: strict depth limit and dynamic maximum tree depth limit. The depth of tree is initially set to be no deeper than the dynamic depth limit. When a tree does not exceed the dynamic maximum depth, it participates in the current population. If a tree is deeper than the dynamic maximum depth but does not exceed the strict maximum depth, it is evaluated by 10-fold cross-validation (CV). In the case that it is the best tree for the respective two-class problem currently, the dynamic maximum depth is increased and the new individual is allowed to join the population; otherwise, the new individual is rejected. If an individual is deeper than the strict maximum depth, it is rejected and one of its parents enters the population. Once the dynamic maximum depth is increased, it would never be lowered again. The initial maximum dynamic depth limit is set to be small so as to force the GP to look for simpler trees firstly before accepting complex solutions.

1

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

The Analysis of Microarray Datasets Using a Genetic Programming

1

The Analysis of Microarray Datasets Using a Genetic Programming

1

1

The Analysis of Microarray Datasets Using a Genetic Programming

2T2Fig 1. The crossover and mutation operators

The standard crossover and mutation operators are deployed here. In the crossover process, two random nodes are chosen from both parent trees, and then the respective branches are swapped so as to create two offspring. The mutation operator randomly chooses a node from the parent tree and replaces it with a new randomly generated sub-tree containing only the features excluded in the parent tree. As the task of the GP is to discover potentially important genes in a huge number of candidates, the mutation operator is important to enhance the exploration in the huge search space.

Sometimes the numbers of samples in different classes are not balanced. To avoid the problem of data skewness problem, a widely deployed method is to only use a small subset of the class whose number of samples is much larger than that of the other class. But as the subset of the data is not

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email