摘要
为了提高基于智能手机的人体行为识别率,优化多分类器集成系统的泛化性能及个体分类器的差异性,提出了基于差异性增量聚类(Diversity Measure Increment-Affinity Propagation clustering,DMI-AP)的选择性集成人体行为识别模型。首先对训练集的所有样本进行bootstrap抽样并训练基分类器,选出大于平均识别率的基分类器构成分类器集合;然后将集合的基分类器作为聚类对象进行分组,通过计算基分类器间的双误差异性值求出表征个体分类器特征的双误差异性增量值,输入近邻传播聚类算法得到k个类簇,选取每簇的中心分类器构成多分类器集成系统;最后使用等概率均值法融合k个分类器的输出结果。实验表明,该模型算法使个体分类器的差异性增大、分类器搜索空间缩小;与传统的Bagging,Adaboost以及RF方法相比,该模型的识别准确率平均提高了8.11%。
To improve the accuracy of human activity recognition based on mobile phone,and optimize the generalization performance of multiple classifiers ensemble system and the diversity of individual classifier,an activity recognition model based on selective ensemble learning of diversity measure increment-affinity propagation clustering(DMI-AP)was proposed.Firstly,all the samples are bootstrapped and base classifiers are trained in the training set.The mode selects the base classifiers whose accuracy is greater than the average accuracy.The classifier set consists of the selected classifiers,and then the base classifiers of the training set are chosen to cluster,the double default diversity increment value are got by calculating the double default diversity measure value between base classifiers.The value is clustered by the affinity propagation clustering algorithm and divided into k clusters.Each cluster's center classifier forms multiclassifier systems.Finally,the outputs of classifiers are fused by calculating the average.The experimental results show that the diversity of individual classifier increases and the searching space of the classifier decreases by using the DMIAP model.Compared with the traditional Bagging,Adaboost and RF methods,the recognition accuracy of the proposed model is improved by 8.11%.
出处
《计算机科学》
CSCD
北大核心
2018年第1期307-312,共6页
Computer Science
基金
国家自然科学基金资助项目(61373116)
陕西省教育科学"十二五"规划课题(SGH140601)
西安邮电大学校青年基金项目(ZL2014-27)
西安邮电大学研究生创新基金项目(103-602080004)资助
关键词
选择性集成
差异性增量
近邻传播聚类
行为识别
Selective ensemble
Diversity measure increment
Affinity propagation clustering
Activity recognition
