主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

中国管理科学 ›› 2026, Vol. 34 ›› Issue (2): 103-119.doi: 10.16381/j.cnki.issn1003-207x.2023.1637cstr: 32146.14.j.cnki.issn1003-207x.2023.1637

• • 上一篇    下一篇

考虑文本结构特征的产品召回监督主题模型及应用研究

施文, 渠玉杰(), 王小双   

  1. 中南大学商学院,湖南 长沙 410083
  • 收稿日期:2023-10-07 修回日期:2024-03-21 出版日期:2026-02-25 发布日期:2026-02-04
  • 通讯作者: 渠玉杰 E-mail:jsqyj3332@csu.edu.cn
  • 基金资助:
    国家自然科学基金项目(72471246);国家自然科学基金项目(72293574);国家自然科学基金项目(71971219);湖南省杰出青年基金项目(2022JJ10084);湖南省教育厅科学研究重点项目(23A0019)

Supervised Topic Modeling and Application Research on Product Recall Considering Textual Structural Features

Wen Shi, Yujie Qu(), Xiaoshuang Wang   

  1. School of Business,Central South University,Changsha 410083,China
  • Received:2023-10-07 Revised:2024-03-21 Online:2026-02-25 Published:2026-02-04
  • Contact: Yujie Qu E-mail:jsqyj3332@csu.edu.cn

摘要:

产品召回是社会重大的公共安全问题,如何能及时地制定正确的召回决策一直是政府和企业工作的目标。海量的在线缺陷信息投诉文本已成为产品召回的重要线索,为解决该问题提供了契机。考虑到缺陷投诉具有短文本并带有辅助信息的特点,本文构建了一个全新的考虑缺陷投诉辅助信息及针对短文本的四层段落监督主题模型(MsSTM),该模型能在大规模的在线缺陷投诉中结合辅助信息提取出潜在的缺陷主题,并据此估计召回发生的概率。利用国内投诉量前20的汽车制造商的166096条缺陷投诉数据和对应的319次召回数据分析发现:MsSTM能从缺陷投诉中提取出有关核心部件、车身附件和售后服务的关键主题,且主题强度与投诉数量、制造商特征等有关联;MsSTM能为不同品牌的制造商识别其缺陷突出程度,其中国内合资品牌的问题主要集中于车内异味、导航系统和悬挂系统,而自主品牌售后服务问题最为突出。最后,MsSTM对于汽车召回的预测稳定性明显优于众多对比模型,ROC-AUC值达到了82.74%,相比于对比模型中最优的ROC-AUC值提升了10.42%。该研究为制造商和政府部门召回管理提供决策依据。

关键词: 产品召回, 在线缺陷投诉, 辅助信息, 段落监督主题模型

Abstract:

Product recalls are a growing concern for public safety, leading governments and businesses to prioritize timely and accurate recall decisions. While recalling defective vehicles is crucial for consumer safety and property rights, it can also have significant financial and reputational consequences for automobile manufacturers. To mitigate these risks, companies are shifting focus towards proactive defect detection and prevention measures. The rise of online defect complaints, facilitated by advances in internet and social media, presents an opportunity to address this issue. However, existing research on product recalls has primarily focused on their causes and consequences, neglecting recall prediction. Similarly, research on defect complaints has centered on categorization, factors influencing complaints, and organizational strategies, often overlooking important auxiliary information within complaint narratives. Thus, there is a need to explore effective approaches for utilizing online defect complaint data to predict automobile recalls.

A novel four-layer metadata-based supervised segmented topic model (abbreviated MsSTM) is proposed, which integrates auxiliary information from defect complaints and addresses the challenges associated with analyzing short texts. This model facilitates the extraction of latent defect topics and estimation of recall probabilities using a large-scale dataset comprising online defect complaints. Experimental findings demonstrate that MsSTM successfully extracts topics about core components, body accessories and after-sales service from defect complaints. Furthermore, the intensity of these topics exhibits correlations with complaint quantity and manufacturer characteristics. The domestic automakers should pay more attention to the after-sales service, while the joint ventures should care more about the odor, navigation and suspension system. Finally, the predictive stability of MsSTM for automobile recalls surpasses that of various comparative models, achieving an impressive ROC-AUC value of 82.74%, reflecting a notable 10.42% enhancement compared to the best ROC-AUC value among the comparative models. The findings carry important implications for relevant automobile manufacturers and government agencies in devising recall strategies, promoting timely enhancements in automotive safety, mitigating the occurrence of traffic accidents, and safeguarding consumer safety and property rights.

Key words: recall prediction, online defect complaints, auxiliary information, supervised segmented topic model

中图分类号: