基于遗传规划算法的特征构建方法研究

Study on the feature construction method based on genetic programming

特征构建(Feature Construction)为了改变分类问题的输入空间,提高分类的性能。遗传规划(Genetic Programming)算法表示方法灵活,适合处理特征构建任务。本研究提出一种基于遗传规划算法的多特征构建方法(gpfcm),此方法利用信息增益比作为遗传规划算法的适应度函数,保存遗传规划算法运行过程中较好的个体作为构建的特征。搜集6个UCI机器学习数据集,选择k最近邻(knn)和决策树(C4.5)作为分类算法,实验比较本研究提出的方法(gpfcm)构建的特征与其它2种方法(gpfcs和gpfcai)构建的特征及与原始特征的特征组合的实验效果。结果表明:本研究的方法构建的特征对大多数数据集能取得比其它方法构建的特征更好或相当的分类效果;利用遗传规划算法构建的高层次的特征能改善分类算法的预测性能。

The purpose of feature construction tasks is to change the input space of classification problem and improve the classification performance.Genetic programming algorithm is suitable for handling feature construction tasks due to its flexible representation.This paper proposes one multiple feature construction method based on genetic programming(gpfcm),which uses the information gain ratio as the fitness function of genetic programming algorithm,preserves excellent individuals during the running process of genetic programming.Six UCI machine learning datasets are collected.K-nearest neighbors and C4.5 decision tree algorithms are selected as classification algorithms.Experiments are done to compare the experimental effect of the feature set constructed by the method(gpfcm)in this paper,the other two feature sets constructed by the other two methods(gpfcs and gpfcai)and these features combined with original features.The experimental results show that the features constructed by our method can obtain better classification effect than the features constructed by other methods,or the features constructed by our method can obtain equivalent performance with the features constructed by other methods.The high-level features constructed by genetic programming algorithm can improve predictive performance of classification algorithms.