Scholarlink

<p>A patent is a valuable intellectual property only when it is granted and held for the long term, and patent grant prediction is a potential strategy for reducing the uncertainty of innovation. Existing machine learning-based prediction models lack interpretability and fail to predict whether granted patents will continue to be maintained, making it difficult to effectively mitigate innovation risks. In this study, we propose an early-stage patent grant prediction model characterized by high interpretability. (1) First, we employ the KAN model for prediction, which replaces traditional neural networks with spline functions, endowing the model with interpretability and the ability to generate formula. (<InternalRef RefID="Equ2">2</InternalRef>) Additionally, we introduced ensemble learning to enhance the performance of the KAN model, resulting in the development of the ENsemble Kolmogorov-Arnold Network (EN-KAN) model. We tested the model on Artificial Intelligence, Biopharmaceuticals, and Electronic Communications datasets and demonstrated strong performance while maintaining high interpretability. (3) Moreover, our study reveals that factors at the examiner-level and the patent-level have the greatest impact on patent grant. However, there are some differences in the influencing factors of patent grant across different fields.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Balancing accuracy and explainability: an ensemble-KAN model for patent grant prediction

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Jing Shi,
  • Xinyi Peng,
  • Xizhen Qiao,
  • Yin Ming,
  • Xiao Liu

摘要

A patent is a valuable intellectual property only when it is granted and held for the long term, and patent grant prediction is a potential strategy for reducing the uncertainty of innovation. Existing machine learning-based prediction models lack interpretability and fail to predict whether granted patents will continue to be maintained, making it difficult to effectively mitigate innovation risks. In this study, we propose an early-stage patent grant prediction model characterized by high interpretability. (1) First, we employ the KAN model for prediction, which replaces traditional neural networks with spline functions, endowing the model with interpretability and the ability to generate formula. (2) Additionally, we introduced ensemble learning to enhance the performance of the KAN model, resulting in the development of the ENsemble Kolmogorov-Arnold Network (EN-KAN) model. We tested the model on Artificial Intelligence, Biopharmaceuticals, and Electronic Communications datasets and demonstrated strong performance while maintaining high interpretability. (3) Moreover, our study reveals that factors at the examiner-level and the patent-level have the greatest impact on patent grant. However, there are some differences in the influencing factors of patent grant across different fields.