Uncovering gold in ash: identifying sleeping beauties among massive unprofitable patents

This paper proposes an innovative deep-learning framework with multi-modal features to determine whether a currently unprofitable patent is a sleeping beauty at an early stage. Patent features include the textual content as well as the networked background information, such as the inventors and assignees, as well as the previous works they have created. The framework uses a Transformer to compare the patent with news or analytical reports concerning technological development trends, mining its content both semantically and syntactically. An active graphical convolutional network, mining the innovation collaboration network of a patent, is also employed as part of the framework to reveal the relationship between patents, companies, and inventors. This framework finally utilizes the obtained features to construct a multi-head self-attention model to predict a patent with the probability of being a sleeping beauty. This paper examines the proposed model by comparing it to several well-known baseline methods using real-world cases from the United States Patent and Trademark Office (USPTO). The proposed deep learning solution outperforms all baseline methods according to all performance metrics. Its long-term forecasting accuracy significantly exceeds its rivals. In the ablation experiments, features extracted from texts and networks are shown to improve the performance of prediction models.