Interpretable patent recommendation with knowledge graph and deep learning.

Area of Science:

• Intellectual Property Management
• Artificial Intelligence
• Data Science

Background:

• Patent transfer is crucial for competitive advantage, but selecting suitable patents is challenging due to large volumes.
• Existing patent recommendation systems often overlook patent quality and lack explainability, hindering effective decision-making.
• The need for interpretable recommendations that incorporate patent quality is critical in the patent transfer context.

Purpose of the Study:

• To propose an interpretable patent recommendation method that addresses the limitations of existing approaches.
• To enhance patent transfer decision-making by integrating patent quality and recommendation explanations.
• To develop a novel method leveraging knowledge graphs and deep learning for improved patent recommendations.

Main Methods:

• Organized heterogeneous patent information into a knowledge graph.
• Extracted connectivity and quality features from the knowledge graph for patent-company pairs.
• Designed an interpretable recommendation model combining deep neural networks and relevance propagation.

Main Results:

• Achieved average precision of 0.596, recall of 0.636, and mean average precision of 0.584.
• Demonstrated performance improvements of 7.28% in precision, 18.35% in recall, and 8.60% in mean average precision over best baselines.
• Successfully interpreted recommendation results by identifying key contributing features.

Conclusions:

• The proposed interpretable patent recommendation method effectively improves patent selection for transfer.
• Integrating knowledge graphs and deep learning enhances recommendation accuracy and provides valuable explanations.
• The method offers a significant advancement for companies navigating complex patent landscapes.