Characterizing Decades of Technological Advances with Graph Neural Networks: An Innovation Network Perspective

Abstract

We leverage the detailed filing information of nearly seven million patents issued by the United States Patent and Trademark Office (USPTO) from 1975 to 2020 and construct a large-scale innovation network based on the forward and backward citations between patents. We employ several state-of-the-art graph neural network algorithms (Node2Vec, Attri2Vec, and GraphSAGE) to extract meaningful representations (embeddings) from patents by taking into account of the innovation network structure. Then, patents are clustered into groups of patents with a strong profile and structural similarity. In conclusion, the predictive power of patents’ learned embeddings demonstrates its usefulness for characterizing the innovation growth over the decades. These embeddings can effectively predict potential linkages between patents that provide a path for future innovation.