西安电子科技大学学报(自然科学版)
西安電子科技大學學報(自然科學版)
서안전자과기대학학보(자연과학판)
JOURNAL OF XIDIAN UNIVERSITY(NATURAL SCIENCE)
2014年
5期
48-53
,共6页
毛莎莎%熊霖%焦李成%张爽%陈博
毛莎莎%熊霖%焦李成%張爽%陳博
모사사%웅림%초리성%장상%진박
集成分类器%旋转森林%支撑矢量机%核匹配追踪
集成分類器%鏇轉森林%支撐矢量機%覈匹配追蹤
집성분류기%선전삼림%지탱시량궤%핵필배추종
classifier ensemble%rotation forest%support vector machine%kernel matching pursuit
为了增强集成系统中各分类器之间的差异性,提出了一种使用旋转森林策略集成两种不同模型分类器的方法,即异构多分类器集成学习算法。首先采用旋转森林对原始样本集进行变换划分,获得新的样本集;然后通过特定比例选择分类精度高的支撑矢量机或分类速度较快的核匹配追踪作为基本的集成个体分类器,并对新样本集进行分类,获得其预测标记;最后结合两种模型下的预测标记。该算法通过结合两种不同分类器模型,实现了精度和速度互补,将二者混合集成后改善了集成系统泛化误差,相比单个模型集成提高了系统分类性能。对 UCI数据集和遥感图像数据集的仿真实验结果表明,文中算法相比单一分类器集成缩短了运行时间,同时提高了系统的分类准确率。
為瞭增彊集成繫統中各分類器之間的差異性,提齣瞭一種使用鏇轉森林策略集成兩種不同模型分類器的方法,即異構多分類器集成學習算法。首先採用鏇轉森林對原始樣本集進行變換劃分,穫得新的樣本集;然後通過特定比例選擇分類精度高的支撐矢量機或分類速度較快的覈匹配追蹤作為基本的集成箇體分類器,併對新樣本集進行分類,穫得其預測標記;最後結閤兩種模型下的預測標記。該算法通過結閤兩種不同分類器模型,實現瞭精度和速度互補,將二者混閤集成後改善瞭集成繫統汎化誤差,相比單箇模型集成提高瞭繫統分類性能。對 UCI數據集和遙感圖像數據集的倣真實驗結果錶明,文中算法相比單一分類器集成縮短瞭運行時間,同時提高瞭繫統的分類準確率。
위료증강집성계통중각분류기지간적차이성,제출료일충사용선전삼림책략집성량충불동모형분류기적방법,즉이구다분류기집성학습산법。수선채용선전삼림대원시양본집진행변환화분,획득신적양본집;연후통과특정비례선택분류정도고적지탱시량궤혹분류속도교쾌적핵필배추종작위기본적집성개체분류기,병대신양본집진행분류,획득기예측표기;최후결합량충모형하적예측표기。해산법통과결합량충불동분류기모형,실현료정도화속도호보,장이자혼합집성후개선료집성계통범화오차,상비단개모형집성제고료계통분류성능。대 UCI수거집화요감도상수거집적방진실험결과표명,문중산법상비단일분류기집성축단료운행시간,동시제고료계통적분류준학솔。
In order to boost the diversity among individual classifiers of an ensemble,a new ensemble method is proposed that combines two different classifier models via a transformation of rotation forest, named by isomerous multiple classifier ensemble.Firstly,the original samples are transformed and divided by the rotating forest to obtain new samples.Then support vector machine with the high accuracy of classification or kernel matching pursuit with the speedy classification is selected as a basic classifier model based on a special proportion,the selected classifier is used to classify the new samples,and the predictive labels are obtained.Finally,the predictive labels given by two different models are combined to obtain the final predictive labels of an ensemble.Particularly,the proposed method achieves the complementarity of accuracy and speed by combining two different classifier models, and it is important that isomerous classifier ensemble improve the generalization error of an ensemble and increases the classification performance.According to the experimental results of classification for UCI datasets and remote sensing image datasets,it is illustrated that the proposed method shortens obviously the running time and improves the accuracy of classification,compared with an ensemble based on the single classifier model.