基于随机校正层级同质性的金融投诉文本聚类参数确定方法

A Parameter Determination Method for Financial Complaint Text Clustering Based on Adjusted Hierarchical Homogeneity

摘要

金融投诉文本聚类对参数敏感，现有内部指标又难以反映业务语义，并可能因簇数增加而高估过碎划分。为此，提出随机校正层级同质性参数确定方法AHH：以文本自带的两级业务标签为弱监督信号，保留V-measure同质性的非对称方向，舍弃限制细分的完整性分量，引入期望互信息校正偶然一致性，并调和粗、细两层得分。CFPB的32组候选实验中，AHH选出的划分在留出NMI-ARI联合指标上排名第一，优于五种基线；在另外八个公开数据集上总体保持前列，OnCoCo揭示了粗层标签退化时的适用边界。结果表明，AHH可将依赖经验的几何试参转化为业务语义驱动、过程可复现的自动选参。

Abstract

Financial complaint text clustering is sensitive to parameter settings. Existing internal indices rarely reflect business semantics and may overrate fragmented partitions as cluster count increases. This paper proposes Adjusted Hierarchical Homogeneity (AHH), which uses two-level business labels as weak supervision, retains the asymmetric direction of V-measure homogeneity, removes the completeness component that discourages legitimate refinement, corrects incidental agreement with expected mutual information, and harmonically aggregates coarse- and fine-level scores. Among 32 CFPB candidates, AHH selects the partition ranked first by the held-out NMI-ARI joint index and outperforms five baselines. It remains competitive on eight additional public datasets, while OnCoCo identifies the applicability boundary caused by a degenerate coarse label. AHH therefore turns experience-based geometric trial and error into reproducible parameter selection driven by business semantics.

关键词

聚类评价指标 / 聚类参数确定 / 弱监督 / 细粒度问题发现 / 双标签

Key words

text clustering / clustering parameter determination / weak supervision / hierarchical homogeneity / fine-grained issue discovery

参考文献

[1] 贺文灏, 吴春江, 周世杰, 等. 融合无监督SimCSE的短文本聚类研究[J]. 计算机科学, 2023, 50(11): 71-76.

[2] 徐菲菲, 陈赛红. 中文文本主题聚类算法研究综述[J]. 上海电力大学学报, 2021, 37(6): 613-619.

[3] 苏桂娜. 基于标准化自适应参数的密度聚类算法研究[J]. 电子技术与软件工程, 2023(3): 247-251.