一种改进的随机森林Boost多标签文本分类算法  

AN IMPROVED RANDOM FOREST BOOST MULTI-LABEL TEXT CLASSIFICATION ALGORITHM

作　　者：邵孟良 齐德昱[2] Shao Mengliang;Qi Deyu(Department of Computer Science,Software Engineering Insitute of Guangzhou,Guangzhou 510990,Guangdong,China;School of Computer Science and Technology,South China University of Technology,Guangzhou 510006,Guangdong,China)

机构地区：[1]广州软件学院计算机系,广东广州510990 [2]华南理工大学计算机科学与工程学院,广东广州510006

出　　处：《计算机应用与软件》2022年第11期215-221,303,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61070015);广东省前沿与关键技术创新项目(2014B010110004);广东省普通高校重点项目(自然)(2019GZDXM020);广州软件学院校级科研团队建设项目(ST202002)。

摘　　要：针对目前Boosting算法计算成本高、学习时间长的问题,提出一种改进的随机森林提升(RF-Boost)算法(IRF-Boost)。对训练特征进行排序;在每个Boosting轮中,过滤并使用排序靠前特征的较小子集;根据权重选择一个特征构建新的弱假设,弱假设搜索空间的大小从k降低至1。实验检验并分析了信息增益、卡方、GSS系数、互信息、优势比、F1得分和准确度共7种特征排序方法。实验结果表明:在所评价的特征排序法中,互信息最适用于RF-Boost;IRF-Boost的效率优于RF-Boost及AdaBost.MH,即IRF-Boost是解决实践应用和专家系统中分类问题的较好选择。The current boosting algorithm has the problem of high computational cost and long learning time,therefore we propose an improved RF-Boost algorithm(IRF-Boost).We sorted the training features,and filtered and used the smaller subsets of the top features in each boosting round.A feature was selected according to the weight to build a new weak hypothesis,and the size of the weak hypothesis search space was reduced from k to 1.Seven feature ranking methods(information gain,chi square,GSS coefficient,mutual information,advantage ratio,F1 score and accuracy)were tested and analyzed.The experimental results show that,mutual information is most suitable for RF-Boost,and the efficiency of IRF-Boost algorithm is better than that of RF-Boost and AdaBost.MH,which means IRF-Boost is a better choice to solve classification problems in practical applications and expert systems.

关 键 词：BOOSTING算法 特征排序 多标签学习 文本分类 弱假设 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]