Abstract
Considering patents tend to be an essential strategic tool for the keen business competitions, patent value assessment help firms identify their important inventions in order to sustain the core competitiveness. In this research, we propose a patent valuability prediction system that can automatically classify a patent into valuable or non-valuable patent. In our proposed patent valuability prediction system, we employ twenty-seven variables extracted from the public available patent information and adopt an ensemble learning algorithm for constructing a patent valuability prediction model. Due to the difficulty of accessing explicit values of patents, proxies are adopted as the training datasets and the questionnaires with true patent values are collected from domain experts as the testing dataset. The proxies employed are both the litigated patents vs. failed-to-maintain patents, and transferred patents vs. failed-to-maintain patents. Moreover, we also combine the litigation and transfer proxies to form the third proxy dataset for training. Accordingly to our evaluation results, the patent valuability prediction system using the litigation proxy as the training dataset exhibits a satisfactory prediction accuracy (i.e., 73.3%), which is higher than that achieved by the use of the transfer proxy or the combined proxy. Moreover, the prediction rate of the use of the litigation proxy can reach as high as 86.7%, with an acceptable recall rate of 76.5%.