Deep learning for predicting patent application outcome: The fusion of text and network embeddings

Patents have been increasingly used as an instrument to study innovation strategies and financial performance of firms recently. Early prediction of patent application success can help firms make better decisions about their investment and innovation strategies. However, predicting patent application outcome is a difficult task that requires the understanding of both deep domain knowledge and complicated legal procedures. In this paper, we propose a novel deep learning framework to mine both the text content and context network, and then fuse these two aspects of features to train a forecasting model to predict the outcome of patent applications. To evaluate the proposed framework, we collect a real-world dataset from the United States Patent and Trademark Office (USPTO). Our method significantly outperforms previous models (e.g., Doc2vec, SciBERT, and PatentBERT) in various metrics, reaching an F1 score of 75.01 percent, and remains robust on different data samples and different scales. Ablation experiments verify that both text and network features help improve the performance of prediction models.