DSM-Forest算法对计算机多类数据学习分类性能的影响

Influence of DSM-Forest algorithm on the learning and classification performance of computer multi-category data

综合运用D-S证据理论和Forest算法,构建得到一种以证据理论为基础的多类半监督分类算法。先以随机方式对多类数据进行组合分类共得到两类数据,再对初始分类器进行训练;利用D-S证据理论来融合部分分类器对未标记样本的分类情况。研究结果表明:随着分类器的数量上升,分类准确率表现为持续升高的现象,但上升幅度不断减小,还会引起学习耗时的快速上升。进行实际半监督学习时,需综合考虑分类准确率与学习耗时来确定最终的分类器数量。当可信度阈值t增大后,可以获得更多的没有参与训练的未标记样本数量。在t逐渐增大的过程中,分类器的分类正确率表现为先上升后下降的现象,并在t等于0.8的条件下到达最大值。

A multi-class semi-supervised classification algorithm based on evidence theory was constructed by using D-S evidence theory and Forest algorithm. Firstly,the data of multiple categories are combined and classified in a random way to get two types of data,and then the initial classifier is trained. D-S evidence theory is used to fuse the classification of unlabeled samples by partial classifiers.The results show that: with the increase of the number of classifiers,the classification accuracy shows a phenomenon of continuous increase,but the increasing range is decreasing,will also cause a rapid increase in learning time. In actual semi-supervised learning,classification accuracy and learning time should be taken into account to determine the final number of classifiers. When the credibility threshold t increases,more unlabeled samples without training can be obtained. In the process of t gradually increasing,the classification accuracy of the classifier shows the phenomenon of first rising and then falling,and reaches the maximum value under the condition of t = 0. 8.