By Liu Xia, Huang Can, Yu Xiaofeng

As the number of patent applications in CNIPA (China National Intellectual Property Administration) increases, patent value is of great interest throughout industries, governments, and academies. However, the existing statistic and econometric models cannot take advantage of huge samples of patent data for value prediction. Based on more than 850,000 patent applications field in 2010 and 2011, this paper provides a machine learning approach to predict patent forward citations at an early stage by using multiple patent indicators that can be defined immediately after the relevant patents are public. The developed model could provide a prediction on whether the relevant patent would receive forward citations; however it was weak in differentiating between high and low citations. Moreover, based on the Gini impurity, features of backward citations provide more information for value prediction. In other words, the prior art search process during the patent examination should be focused on. Finally, the paper discusses the limitations of the adopted model, as well as improvement methods for further studies.

This article is published in Journal of The China Society for Scientific and Technical Information, 2019, 38(04): 402-410. (in Chinese)