一种随机平均分布的集成学习方法

AN ENSEMBLE LEARNING METHOD WITH RANDOM EVENLY DISTRIBUTION

像传统机器学习一样,样本的不平衡分布会影响深度学习分类器的预测能力,在语音情感识别环境下,情感数据的不平衡分布是一种常态。基于卷积循环神经网络和注意力模型,提出一种随机平均分布的集成学习方法(Redagging),用来消除样本的不平衡分布。Redagging按照机会均等原则,等概率地把训练样例随机放入子训练样本,通过降低样例重复率提升基分类器的性能,进而增强综合分类器的预测能力。在IEMOCAP和EMODB情感数据库的实验表明,从未加权平均召回率和F1值两个方面,Redagging都优于Bagging和其他不平衡学习方法,验证了其有效性。

Like traditional machine learning,the imbalance issue will have a negative impact on the predictive ability of deep learning classifiers,and imbalanced distribution is a normal phenomenon in the context of speech emotion recognition.Based on convolutional recurrent neural network and attention model,a random evenly distributed aggregation method(Redagging)which aims at dealing with imbalance problem is proposed.According to the principle of equal opportunity,Redagging randomly and uniformly put training examples into sub samples.It improved the performance of the base learners by reducing duplicate examples,and then enhanced the predictive ability of the final learner.The experiments on the IEMOCAP and EMODB sentimental databases show that Redagging is better than Bagging and other imbalance learning methods in terms of UAR and F1 score,proving its effectiveness.