集成学习通过训练多个不同的个体学习器并加以结合，从而获得比单一学习器更好的学习效果和泛化能力，已在机器学习领域获得广泛关注。集成剪枝技术可以从原始的集成系统中挖掘出一个更合适的学习器子集，从而使得集成系统的学习效果更好的同时系统规模更小。目前已经有多种集成剪枝技术应用于集成学习系统的创建过程中，这些剪枝技术很多是基于个体学习器的准确性或者多样性进行选择的。Kappa-Error图将学习器的准确性和多样性放在一张点云图中统一进行衡量，在一般的研究中Kappa-Error图多是用来对分类器之间的多样性和准确性进行分析，但实际上图中包含的多样性和准确性信息是用于集成剪枝处理的有效工具。因此，本文拟基于对Kappa-Error图的分析设计出一种新的集成剪枝方案。 在传统的Kappa-Error图中，位于左下角的点代表具有较高多样性和准确性的分类器对，可以用一条斜率为负的直线将位于左下方的点分割开来，以此作为剪枝后的分类器子集。不过，图中的点表示的是一对分类器，在选择过程中需要考虑两个分类器的剪枝顺序，而且很有可能会选择重复的分类器个体，也很有可能选择那些较差准确率的分类器。因此本文提出的集成剪枝方案对传统的Kappa-Error图进行了改进，改进后Kappa-Error图中的点代表一个分类器，纵坐标表示该分类器的准确率，横坐标表示该分类器和其他所有个体分类器的Kappa值的平均。这种改进保留了传统Kappa-Error图的多样性和准确性信息，并且使得原始图形无法用于剪枝的几点因素得到改善，从而可以方便的用于集成剪枝处理。 新算法的关键在于寻找出合适的直线（y=k*x+b），即k和b的值。实际上，当k确定时，b的取值范围可以根据点云的位置计算出来。因此，新算法只需对k进行设置。为寻找最优参数，本文提出一种遍历直线夹角的方法，通过遍历直线的夹角来控制斜率k的取值变化，进而计算b的范围，最后以集成系统的分类准确率作为度量标准来判断参数的最优解。 为了验证该剪枝算法的有效性，本文将该剪枝方案应用到Bagging集成中，利用UCI上不同领域的数据集进行分析和对比。实验结果表明，新的剪枝方法能很好地提高集成系统的整体性能，同时减小集成系统的规模。 之后，本文将该集成算法应用到基于语音信号的抑郁识别研究中。实验通过不同言语方式和情绪效价下的刺激材料采集了大量语音数据，然后对不同言语方式和情绪效价的语音进行了分析和比较，它们之间的差异性为集成学习系统的应用提供了良好的基础。集成系统构建的第一步是个体分类器的构建，研究将该领域表现较好的支持向量机作为个体分类器，训练的过程是在不同的语音数据上进行的，并且每次训练都选择不同的语音特征和模型参数，目的在于增加个体分类器的准确率和它们之间的多样性。第二步，基于Kappa-Error图的剪枝过程被应用到该系统中，剪枝后的分类器按照并行结构集成起来并采用简单的多数投票法来进行最后的决策。实验结果显示，本文提出的集成方案能显著提高抑郁识别的分类准确率，并且基于Kappa-Error图的剪枝方案明显优于其他剪枝策略。 综上，本文的研究内容主要有两点：（1）本文基于Kappa-Error图设计了新的集成剪枝方案，并在UCI的多个数据集上验证了该剪枝方案的有效性；（2）本文将基于该剪枝方案的Bagging集成系统应用到语音抑郁识别中，从而显著提高了语音抑郁识别的准确率。

Ensemble learning has a much higher learning efficiency and generalization error than unitary learning by training and combining individual learner. It has been widely concerned and studied in the machine learning area. The ensemble pruning technology make it possible to select a more suitable congregation from the original ensemble learning system, to improve the performance and decrease the scale of system. Currently, these technologies have been widely applied in the establishment of an ensemble learning system, while most of them select on the basis of accuracy and diversity of individual learner. The Kappa-Error Figure measures the accuracy and diversity of individual learner in one cloud chart, which can help us to pruning. In this article, a new ensemble pruning scheme which designed on the basis of Kappa-Error Figure is introduced. Then, the new system is employed in detecting depression in speech to raise the accuracy of depression recognition. In a classic Kappa-Error Figure, points positioned in the lower left corner presented pairs of classifiers with high accuracy and diversity. We can use a straight line with a negative slope to separate points located in the lower left corner, which are integrated to be the ensemble learning system. However, in that case, the pruned order of the two classifiers is fuzzy. Moreover, duplicate classifier and low-accuracy classifier may be selected as well. In order to solve these problems, classic Kappa-Error Figure has been improved. The new one not only reserved the classic one’s feature of accuracy and diversity, but also made it to be used in pruning easily. The key of the algorithm is to find the right straight line (y = k * x + b), in other words, the parameter k and b. Practically, the parameter b can be determined by the point on the cloud chart when k is acknowledged, so the key of the algorithm is just to set up the value of k. Our study suggested discovered a priority method to seek the value of k by ergotising linear angle. In this way, the range of k can be determined by ergotising the angle, then the optimal number of parameter k and b can be determined by measuring the classification accuracy rate of ensemble learning system. In this research, the new pruning scheme was applied in the Bagging system. The validity of this scheme was further proved by data gained from UCI of diverse fields. Consequently, the pruning scheme is an efficient way to raise the running performance and decrease the scale of the ensemble learning system. Then, the new ensemble algorithm was employed in the speech recognition of depression field. Speech data achieved from stimulating materials under different verbal and emotional valence were analyzed and compared. The difference of the speech data built a good foundation for the application of ensemble learning system. At first, supporting vector machine, which was popular in this filed, was used to construct individual learner. The training process was carried out on different voice data and different vocal features and model parameters were selected in each training. By doing so, the accuracy rate and diversity of systems were increased. Then, the pruning process on the basis of Kappa-Error Figure was employed in the system. Finally, the pruned classier then integrated due to parallel structure in a simple majority voting method manner. It can be clearly seen from the result that, the classification accuracy rate of depression recognition had an obviously rise under the ensemble pruning scheme. In summary, there are two key points in this article. First, a designation of ensemble pruning scheme on the basis of Kappa-Error Figure, and the vitality of the scheme examined by the data sets gained from UCI. Second, the Bagging system based on the pruning scheme is applied to depression recognition in speech, thus significantly improving the accuracy of depression recognition.