基于深度双向分类器链的多标签新闻分类算法

在多标签新闻分类问题中，针对传统分类器链算法难以确定标签依赖顺序、集成模型运行效率低和无法应用复杂模型作为基分类器的问题，提出基于深度神经网络的双向分类器链算法. 该方法利用正向分类器链获取每个标签和前面所有标签的依赖关系，引入逆向分类器链，从正向链最后一个基分类器的输出开始反向学习每个标签和所有其他标签的相关性. 为了提取非线性标签相关性和提高预测性能，使用深度神经网络作为基分类器. 结合2条分类器链的均方误差，使用随机梯度下降算法对目标函数进行有效优化. 在多标签新闻分类数据集RCV1-v2上，将所提算法与当前主流的分类器链算法和其他多标签分类算法进行对比和分析. 实验结果表明，利用深度双向分类器链算法能够有效提升预测性能.

Multi-label news classification algorithm based on deep bi-directional classifier chains

A deep neural network based bi-directional classifier chains algorithm was proposed for multi-label news classification tasks to deal with problems faced by traditional classifier chains method, i.e. hard to determine the order of label dependencies, the inefficiency of integrated models and incapable of using complicated base classifiers. In the proposed method, a forward classifier chain is utilized to obtain the correlation between each label and all previous labels, and a backward classifier chain is involved, starting from the output of the last base classifier in the forward classifier chain, to learn the correlations between each label and all other labels. The deep neural network is employed as a base classifier in order to explore the non-linear label correlation and improve the predictive performance. Br integrating the mean square loss of the two classifier chains, the objective function is optimized by stochastic gradient descent algorithm. The experimental results of the proposed method for multi-label news classification dataset RCV1-v2 were compared with those of current classifier chains methods and other multi-label algorithms. Results show that the deep bi-directional classifier chains can significantly improve the predictive performance.