Domain adaptation based transfer learning for patent transfer prediction

Weidong Liu, Yiming Wang, Keqin Gan, Xiangfeng Luo, Yu Zhang, Cuicui Jiang

Abstract: As intellectual property competition intensifies, significant differences in patent data distribution arise across countries. Addressing cross-national prediction disparities through transfer learning, specifically from patent-developed to patent-developing countries, to enhance patent transfer prediction accuracy, remains challenging. This study addresses two fundamental challenges: (1) how to choose the transfer learning direction, and (2) how to improve the accuracy of patent transfer prediction. We herein propose a domain adaptation transfer learning-based patent transfer prediction model. For the choice of transfer learning direction, we identify key indicators for directional selection by evaluating the influence of patent data indicators and establish criteria for determining the transfer learning direction. For domain adaptation learning, we use the BERT pretrained model, incorporating a lambda layer to extract patent features. A gradient reversal layer is added to narrow the disparity in patent feature distribution between the source and target domains, allowing us to learn domain-shared features. The experiments show that patent quantity and patent transfer speed are important indicators for choosing the transfer learning direction. Additionally, the model significantly improves precision, recall, and F1-score, with improvement rates of 6.85%, 6.83%, and 7.11%, respectively, compared with baseline models.