Core patent forecasting based on graph neural networks with an application in stock markets

Abstract

ABSTRACT Core patents are the most important in a specific technological field. Forecasting core patents is crucial to understanding the development of technology trends. Traditional approaches are based on structured/unstructured features or patent relationships. However, most previous methods focused on specific aspects of patents. We propose a novel framework based on the potential relationships to forecast core patents. An event study approach is applied to analyze the short-term impact of core patents’ granting events on the stock market. We apply three methods to build framework: a baseline model using traditional features and two graph embeddings. Finally, a classification model is used to predict and compare the effectiveness of different inputs: the traditional patent feature index, and the feature vectors output by the Node2vec and GraphSAGE models. We verify the method with data from the communications and biomedical industries and investigate its application to the stock market. Results demonstrate that the graph-embedding features based on the network are superior to traditional patent features. The graph neural network effectively fuses the two sets of information, and forecasting is improved in all models. And we find that the cumulative abnormal returns from core patents’ granting events outweigh those for non-core (ordinary) patents.