基于BR和GBDT的电力信息通信客服系统多标签文本分类

现有电力信息通信(ICT)客服系统主要依靠客服坐席员经验,根据电力ICT系统用户报修信息进行故障类型分类判别,存在在线处理及时性较差、准确性不足的问题.针对上述问题,提出了一种基于集成学习的电力ICT客服系统文本数据的多标签文本分类方法,实现对电力ICT系统的复杂故障类型进行自动化、高准确率分类识别.首先,针对电力ICT系统故障类型识别准确率偏低且低效的问题,提出了基于二元相关性(BR)和梯度提升决策树(GBDT)集成学习的多标签分类方法,将BR和GBDT有机结合实现自动化、高准确率的故障多标签分类.其次,针对电力ICT客服文本数据的多标签分类训练集难以获取的问题,提出一种面向电力ICT客服文本数据的多标签训练集自动化构建方法,实现了高效的电力ICT客服文本多标签分类.实验表明,BR-GBDT方法可以高效处理电力ICT系统复杂故障类型的多标签分类任务,分类性能也优于BR+逻辑回归(LR)和多标签k最近邻(ML-kNN)等典型的集成学习多标签分类方法.

Multi-label Text Classification for Power ICT Custom Service System Based on Binary Relevance and Gradient Boosting Decision Tree

The current power information and communication technology (ICT) custom service system mainly relies on the personal experience of custom service staff to judge fault types according to faults reported by ICT system users. It has the problems of processing untimeliness and inaccuracy. To solve these problems, an multi-label text classification method based on ensemble learning for the power ICT custom service system is proposed which can automatically and accurately identifies the complex fault types of power ICT systems. First, for the problem of low accuracy and efficiency when identifying ICT fault types, an ensemble method based on binary relevance-gradient boosting decision tree (BR-GBDT) is proposed for multi-label text classification, which combines binary relevance and gradient boosting decision tree to improve the classification accuracy. Second, to solve the difficulty in the construction of the multi-label classification training set of power ICT custom service data, an automatic approach is presented to construct the training set of power ICT custom service text, so that efficient classification is realized for power ICT custom service text. Experiment results show that the BR-GBDT method can not only efficiently handle the multi-label classification of power ICT custom service system faults, but also has a better performance than other typical multi-label classification methods, such as BR+logistic regression (LR) and multi-label k-nearest neighbor (ML-kNN).