一种基于决策森林的单调分类方法

发布时间：2018-07-10 06:18

本文选题：单调分类 + 决策树　；参考：《计算机研究与发展》2017年07期

【摘要】：单调分类问题是特征与类别之间带有单调性约束的有序分类问题.对于符号数据的单调分类问题已有较好的方法,但对于数值数据,现有的方法分类精度和运行效率有限.提出一种基于决策森林的单调分类方法(monotonic classification method based on decision forest,MCDF),设计采样策略来构造决策树,可以保持数据子集与原数据集分布一致,并通过样本权重避免非单调数据的影响,在保持较高分类精度的同时有效提高了运行效率,同时这种策略可以自动确定决策森林中决策树的个数.在决策森林进行分类时,给出了决策冲突时的解决方法.提出的方法既可以处理符号数据,也可以处理数值数据.在人造数据集、UCI及真实数据集上的实验数据表明:该方法可以提高单调分类性能和运行效率,缩短分类规则的长度,解决数据集规模较大的单调分类问题.
[Abstract]:Monotone classification is an ordered classification problem with monotonicity constraints between features and categories. There are good methods for monotone classification of symbolic data, but for numerical data, the classification accuracy and operational efficiency of the existing methods are limited. A monotone classification method based on decision forest (monotonic classification method based on decision forestMCDF) is proposed. Sampling strategy is designed to construct decision tree. It can keep the distribution of data subset consistent with the original data set, and avoid the influence of non-monotone data through sample weight. This strategy can automatically determine the number of decision trees in the decision-making forest. In the process of classification of decision-making forest, the solution of decision-making conflict is given. The proposed method can deal with both symbolic and numerical data. The experimental data on UCI and real data sets show that this method can improve the performance and efficiency of monotone classification, shorten the length of classification rules, and solve the problem of monotone classification with large scale of data sets.
【作者单位】：山西大学计算机与信息技术学院;计算智能与中文信息处理教育部重点实验室(山西大学);山西财经大学应用数学学院;
【基金】：国家自然科学基金项目(61673249,61503229) 山西省回国留学人员科研基金项目(2016-004) 山西省研究生教育创新项目(2016BY003)~~
【分类号】：TP181

【参考文献】