| 张广发* **,陈加乐* **,方金云*.基于权力信号的跨表格迁移学习方法研究[J].高技术通讯(中文),2025,35(5):451~460 |
| 基于权力信号的跨表格迁移学习方法研究 |
| Research on cross-tabular transfer learning methods based on power signals |
| |
| DOI：10.3772/j.issn.1002-0470.2025.05.001 |
| 中文关键词: 大数据监督； 政务数据； 权力信号； 表格学习； 迁移学习 |
| 英文关键词: big data supervision, administrative data, power signal, tabular learning, transfer learning |
| 基金项目: |
| 作者 | 单位 | | 张广发* ** | （*中国科学院计算技术研究所北京 100190）
（**中国科学院大学北京 100190） | | 陈加乐* ** | | | 方金云* | |
|
| 摘要点击次数: 55 |
| 全文下载次数: 114 |
| 中文摘要: |
| 为了有效监督和审计政府行使公权力，本文提出了一种基于权力信号的跨表格迁移学习方法，目的是从政务信息系统的表格数据（简称政务表格数据）中自动检测出权力滥用问题。权力信号是公权力行使过程中的关键特征，由关键人、决策、资金、项目和物资5个要素构成。这些权力信号分布在不同的政务表格数据中，政务表格数据结构多样，对权力信号跨表格学习带来挑战。本文设计了一种基于权力信号的跨表格迁移学习框架PowerTab（power tabular transformer），旨在引导模型在政务表格数据上学习通用的权力信号表征，并使用迁移学习将其应用到目标任务的检测模型中。该框架实现了一种在政务表格数据中提取词元级权力特征的方法，使得检测模型具有零样本学习能力。在5个数据集上的实验结果表明本文方法优于基线方法，为政务表格数据的大数据监督提供了一种有效的手段。 |
| 英文摘要: |
| In order to effectively supervise and audit the government’s exercise of public power, this paper proposes a cross-table transfer learning method based on power signals, which aims to automatically detect power abuse problems from the table data of government information systems (referred to as government table data). Power signals are the key features of the process of public power exercise, which are composed of five elements: key person, decision, fund, project, and material. These power signals are distributed in different government table data, and the diversity of government table data structure brings challenges to the cross-table learning of power signals. This paper designs a cross-table transfer learning framework based on power signals, namely PowerTab (power tabular transformer), which aims to guide the model to learn general power signal representations on government table data, and apply them to the detection model of target task by using transfer learning. The framework implements a method to extract token-wise power features in government table data, which endows the detection model with zero-shot learning ability. The experimental results on five datasets show that the proposed method outperforms the baseline methods, and provides an effective means for big data supervision of government table data. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
