面向动态数据流的三支多粒度主题建模方法  

Three-way Multi-granularity Topic Modeling Method Under Dynamic Data Stream

作　　者：李昱洁 杨新 李妙妙[1,2] 李天瑞 LI Yujie;YANG Xin;LI Miaomiao;LI Tianrui(Department of Artificial Intelligence,School of Economic Information Engineering,Southwestern University of Finance and Economics,Chengdu 611130,China;Financial Intelligence and Financial Engineering Key Laboratory of Sichuan Province,Southwestern University of Finance and Economics,Chengdu 611130,China;School of Computing and Artificial Intelligence,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南财经大学经济信息工程学院人工智能系,四川成都611130 [2]西南财经大学金融智能与金融工程四川省重点实验室,四川成都611130 [3]西南交通大学计算机与人工智能学院,四川成都611756

出　　处：《山西大学学报（自然科学版）》2022年第1期31-40,共10页Journal of Shanxi University(Natural Science Edition)

基　　金：国家重点研发计划课题(2019YFB210180);国家自然科学基金(61773324,61876157);教育部人文社会科学青年基金(20YJC630191)。

摘　　要：传统机器学习模型大多依赖封闭世界假设,难以持续挖掘开放环境下动态文本数据流的语义信息。文章考虑动态数据流的粒特征变化,提出了一种基于三支多粒度学习思想的主题建模方法(Three-way Multi-granularity Topic Model,3WMTM)。根据“文本-主题-词汇”的粒度参数演化机理,构建了主题模型的动态多粒度分析框架,在主题不断新增背景下自适应更新阈值。实验结果表明,3WMTM算法在动态环境下能够维持较好的决策精度,即统计上无显著差异;并且3WMTM算法能够在每个决策阶段立即给出判断,因此在决策成本上优于现有模型。Most traditional machine learning models rely on closed-world assumption,which makes it difficult to continuously mine the semantic information of dynamic data in open-world.This paper introduced a topic modeling method(Three-way Multi-granularity Topic Model,3WMTM)considering the variation of granularity characteristics of dynamic data as well.According to the granularity parameter evolution mechanism of"text-topic-vocabulary",a dynamic three-way decision multi-granularity framework of topic modeling is constructed,and the thresholds are updated adaptively.The experimental results show that the 3WMTM algorithm can maintain decision precision in dynamic environment,and is better than the existing models in decision cost.

关 键 词：开放动态环境 三支多粒度学习 主题模型 主题新增 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]