一种自适应的tri-training半监督算法.pdfVIP

计 算 机 系 统 应 用 2016 年 第 25 卷 第 8 期 一种自适应的Tri-Training 半监督算法① 1 2 彭雅琴 , 宫宁生 1(三江学院 计算机科学与工程学院, 南京 210012) 2(南京工业大学 计算机科学与工程学院, 南京 210009) 摘 要: Tri-Training 算法是半监督算法的一种, 在学习过程中容易错误标注无标记样本, 从而降低分类性能, 为 此提出一种 ADP-Tri-Training(Adaptive Tri-Training)算法, 改进协同工作方式, 根据几何中心设置分类器组成, 然 后应用模糊数学理论将多个独立的分类器组合, 使得算法可以在多因素下综合评价样本, 并在此基础上引入遗 传算法动态设置组合权重以适应于具体的样本集, 从而尽可能降低样本标注的错误率, 多个实验结果表明 ADP-Tri-Training 算法具有更好的分类性能. 关键词: Tri-Training 算法; 自适应;遗传算法;差异性度量;半监督 Adaptive Tri-Training Semi-Supervised Algorithm 1 2 PENG Ya-Qin , GONG Ning-Sheng 1(Department of Computer Science and Engineering, Sanjiang University, Nanjing 210012, China) 2(Department of Computer Science and Engineering, Nanjing University of Technology, Nanjing 210009, China) Abstract: Tri-Training algorithm belongs to semi-supervised algorithm,unlabeled samples are often labeled incorrectly in study, and the performance is affected. So the ADP-Tri-Training (Adaptive Tri-Training) algorithm is proposed, cooperative work mode is improved, a classification composition scheme based on geometric center is used, the fuzzy mathematics theory is applied to combine the classifiers, so the algorithm can evaluate the samples by multiple factors, genetic algorithm is introduced to dynamically set the combined weight in order to adapt different sample sets, also it can reduce the error of classifies as far as possible, finally the experimental results show that the proposed algorithm is more effective. Key words: Tri-Training algorithm; adaptive; genetic algorithm; novel diversity measure; semi-supervised 在实际应用中, 有标记样本获取相对困难, 更多 习思想将图中结点对应为样本, 边对应为相似度,