Transfer learning based standard-essential patent prediction with prior transfer direction learning

Abstract

Standard-Essential Patent Prediction (SEPP) holds strategic significance for technological development and international market competition. Traditional SEPP models learned from Standard-Essential Patents (SEPs) with country-specific distribution differences result in different prediction accuracy. Therefore, we propose two propositions: (1) Can transfer learning be leveraged to improve prediction performance of lower-accuracy countries. (2) Can different transfer directions achieve different transfer learning performances. To address these, we propose a transfer learning based SEPP ewith prior transfer direction learning (TLSEPP-PTDL) model. The model uses a mixed transfer learning method, achieving an average accuracy of 92.03 % on four datasets, surpassing the state-of-the-art (SOTA) by 2.03 % and improving precision, recall, and F1-score by 4.25 %, 0.33 %, and 2.25 %, respectively. Moreover, we conduct experiments across countries with different patent volume, standardization rate, and standardization speed, resulting in positive transfer when transfer learning uses source domains with (1) high volume, high rate, and high speed; (2) high volume, low rate, and high speed; (3) low volume, high rate, and high speed.