Interpretable patent recommendation with knowledge graph and deep learning

Abstract

Patent transfer is a common practice for companies to obtain competitive advantages. However, they encounter the difficulty of selecting suitable patents because the number of patents is increasingly large. Many patent recommendation methods have been proposed to ease the difficulty, but they ignore patent quality and cannot explain why certain patents are recommended. Patent quality and recommendation explanations affect companies’ decision-making in the patent transfer context. Failing to consider them in the recommendation process leads to less effective recommendation results. To fill these gaps, this paper proposes an interpretable patent recommendation method based on knowledge graph and deep learning. The proposed method organizes heterogeneous patent information as a knowledge graph. Then it extracts connectivity and quality features from the knowledge graph for pairs of patents and companies. The former features indicate the relevance of the pairs while the latter features reflect the quality of the patents. Based on the features, we design an interpretable recommendation model by combining a deep neural network with a relevance propagation technique. We conduct experiments with real-world data to evaluate the proposed method. Recommendation lists with varying lengths show that the average precision, recall, and mean average precision of the proposed method are 0.596, 0.636, and 0.584, which improve corresponding performance of best baselines by 7.28%, 18.35%, and 8.60%, respectively. Besides, our method interprets recommendation results by identifying important features leading to the results.