关键词: 深度学习, 不均衡, 文本, 分类方法, 深度信念网络, 文档概率, 预训练, 对比算法

Abstract:

In unbalanced text classification， the classification results tend to the majority and ignore the minority， which leads to poor classification effect. The unbalanced text classification method based on deep learning is studied. DA method is used to select unbalanced text features. DA method sets the scoring standard to the minimum value of the difference of document probability correlation， so that the selected text features are evenly distributed in most classes and a few classes to improve the balance of text features. The subset obtained by feature selection is used as the input of the depth belief network composed of multiple constrained Boltzmann machines. The constrained Boltzmann machine obtains the optimal probability distribution of training samples through pre training. The weight of the constrained Boltzmann machine is determined by contrast bifurcation algorithm. After the parameters of the constrained Boltzmann machine are set， the greedy algorithm is used to train the constrained Boltzmann machine iteratively until the whole process is completed text classification. Experimental results show that this method can effectively classify unbalanced text， and the classification accuracy is more than 99.5%.

Key words: deep learning, imbalance, text, classification method, deep belief network, document probability, pre-training, contrast divergence algorithm