面向不均衡小样本训练集的改进Boosting算法.pdfVIP

第4 0卷第2期 Vol. 40, No. 2 2 0 1 0 年2 月 JO URNAL O F UN IVERSITY OF SC IENC E AN D TEC HN OLO GY O F C HINA Feb. 2 0 1 0 文章编号:025322778(2010) 022 01462 06 Boosting 1 1 1, 2 2 程有龙 , 庄连生, 李 斌 , 庄镇泉 ( 1. 微软2教育部多媒体计算与通信联合实验室, 安徽合肥230027; 2. 中国科学技术大学电子科学与技术系, 安徽合肥 230027) :Boosti g . , Boosti g , , , , , . , , , , , booti g . :Boosti g; ;;; ; : TP18 :A doi: 10. 3969/j. iss . 025322778. 2010. 02. 007 Improved Boosting algorithm for small and imbalanced training sets 1 1 1, 2 2 CHENG Youlo g , ZHUANG Lia she g , LI Bi , ZHUANG Zhe qua ( 1. MOE2Microsof t Key Laboratory of Multimedia Computing and Communication, H ef ei 230027, China; 2. Dep artment of Electronic cience and Technology , University of cience and Technology of China , H ef ei 230027, China) Abstract: Traditio al Boosti g algorithms te d to overfit a d be biased towards the major ity class o small a d imbala ced trai i g sets. To addr ess this issue, a improved Boosti g lear i g algorithm with adaptive sample i jecti g a d feature k ock out was proposed. I the tr ai i g process, sy thetic samples were appe ded to the origi al trai i g set to rebala ce it a d disturb a d e ha ce its ge eralizatio ability. The method was tested o both two2class a d multi2class image classificatio problems. Experime t results show that whe the umber of trai i g samples is small, a d the distributio of trai i g set is imbala ced, the proposed method ca e ha ce the ge er alizatio performa ce of Boosti g algorithms effectively. Key words: Boosti g; small sample set; imbala ced trai i g set; feature k ock out; disturbi g; ge e