一句话：TixFusion 用无监督聚类加 LLM 抽取判别信息聚合重复工单，降低标注成本并将缺陷分派提效 3.78 倍。
移动操作系统因复杂性与频繁更新而缺陷频发，厂商常邀请 beta 用户测试并提交工单，但海量工单在 triage（分派）阶段严重拖慢工程师把工单转交给合适开发团队的效率。将指向同一缺陷的重复工单聚合后可批量分派，从而提速，但现有工单聚合方法需要大量标注数据训练，人工成本高、难以在生产环境落地。为此作者提出 TixFusion：一个 LLM 增强的工单聚合框架，采用无监督聚类完成聚合以尽量减少标注数据需求，并引入 LLM 从工单中抽取具有判别力的信息以提升聚合准确率。基于全球顶级移动操作系统厂商华为生产环境数据集的实验表明，TixFusion 在保持低成本的同时优于现有方法；该系统已在华为部署三个多月，处理超过 20 万条工单，使 triage 工程师的处理速度提升 3.78 倍。
Mobile OS often exhibit defects due to their complexity and frequent updates, adversely affecting user experience. To resolve these defects, OS providers enlist beta users to test the OS. During testing, any defects encountered are recorded in tickets and reported back to the providers. However, the sheer volume of these tickets significantly challenges the efficiency of the ticket handling system, especially during the triage phase, where engineers assign tickets to appropriate development teams for resolution. Enhancing triage efficiency is possible by aggregating duplicate tickets related to the same defect, allowing simultaneous assignment of multiple tickets. Nonetheless, current ticket aggregation methods demand substantial labeled data for training, imposing a labor cost that hinders their practical implementation in production environments. To reduce this labor cost, we propose TixFusion, an LLM-augmented ticket aggregation framework. TixFusion employs unsupervised clustering to aggregate tickets, minimizing the need for labeled data, and integrates an LLM to extract discriminative information from tickets, improving aggregation accuracy. Extensive experiments using a dataset collected from the production environment of a top-tier global mobile OS provider Huawei demonstrate that TixFusion outperforms existing methods while maintaining a low labor cost. Additionally, TixFusion has been deployed in Huawei for over three months, processing more than 200,000 tickets, and has increased the processing speed of triage engineers by 3.78 times.
@inproceedings{SunHWZZSZGLWP25,
title = {LLM-Augmented Ticket Aggregation for Low-cost Mobile OS Defect Resolution},
author = {Yongqian Sun and Bowen Hao and Xiaotian Wang and Chenyu Zhao and Yongxin Zhao and Binpeng Shi and Shenglin Zhang and Qiao Ge and Wenhu Li and Hua Wei and Dan Pei},
booktitle = {FSE},
year = {2025},
doi = {10.1145/3696630.3728547}
}