基于多类分类的主动学习改进算法

发布时间：2018-08-30 12:05

【摘要】：对于监督学习模型,足够多的已标注样例是获得高精度分类器的前提条件。但现实中,未标注样例在总样例中占比较大,而人工标注样例成本昂贵,需要控制训练样例集的数量和质量。如何高效地选出具有高分类贡献度的无类标样例进行标注并补充到已有训练集中,以便逐步提高分类器精度与鲁棒性是主动学习亟待解决的关键问题。另外,大部分主动学习研究局限于封闭的样例集,如何将主动学习应用到开放的实际生产环境中且能达到较高的分类精度也非常值得研究。针对BvSB样例选择算法存在的类间均衡性问题和孤立点问题,采用不确定性和代表性相结合的思想,提出了Center+reBvSB样例选择算法。首先利用K-Means聚类选择代表性训练集A,再采用reBvSB样例选择算法挑选代表性边缘均衡样例集合B,最后整合A和B并更新训练集。实验结果表明,该算法选出的样例可以帮助分类器提高分类精度和鲁棒性。将Center+reBvSB样例选择算法集成到主动学习算法中,提出了改进BvSB的主动学习算法。利用在线识别产生的错分样例池结合原始样例池,对分类器进行重新训练,可以进一步提升分类器的识别能力。利用Mnist数据集模拟实际在线识别场景,实验结果表明,改进的主动学习算法具有较好的鲁棒性和较高的识别精度。
[Abstract]:For supervised learning model, sufficient number of labeled samples is a prerequisite for obtaining high accuracy classifiers. However, in reality, unlabeled samples account for a large proportion of the total samples, while manual labeling is expensive, so it is necessary to control the quantity and quality of training samples. It is a key problem to solve the problem of active learning that how to efficiently select the non-class sample with high classification contribution and add to the existing training set in order to improve the accuracy and robustness of classifier step by step. In addition, most of the active learning research is confined to the closed sample set, how to apply active learning to the open production environment and achieve high classification accuracy is also worth studying. Aiming at the problem of inter-class equilibrium and outliers in BvSB sample selection algorithm, a Center reBvSB sample selection algorithm is proposed by combining uncertainty and representativeness. Firstly, K-Means clustering is used to select representative training set A, then reBvSB sample selection algorithm is used to select representative edge equilibrium sample set B. finally, A and B are integrated and the training set is updated. Experimental results show that the proposed algorithm can improve the accuracy and robustness of the classifier. The Center reBvSB sample selection algorithm is integrated into the active learning algorithm, and an improved BvSB active learning algorithm is proposed. The recognition ability of the classifier can be further improved by retraining the classifier by using the error sample pool generated by on-line recognition combined with the original sample pool. The Mnist dataset is used to simulate the real online recognition scene. The experimental results show that the improved active learning algorithm has better robustness and higher recognition accuracy.
【学位授予单位】：河北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP181

【参考文献】