PTNS: patent citation trajectory prediction based on temporal network snapshots.

Area of Science:

• Intellectual Property Management
• Data Science and Machine Learning
• Innovation Studies

Background:

• The knowledge economy necessitates strategic identification of high-value patents for technological innovation.
• Patent citation behavior is complex and exhibits temporal variations, making accurate prediction challenging.
• Existing methods struggle to effectively capture the dynamic nature of patent citation networks.

Purpose of the Study:

• To develop a novel model for predicting patent citation trajectories and identifying impactful patents.
• To effectively model the temporal evolution and complex relationships within patent citation networks.
• To capture phenomena like patent aging and the 'sleeping beauty' effect in citation patterns.

Main Methods:

• Proposed a patent citation trajectory prediction model (PTNS) utilizing temporal network snapshots.
• Employed Relational Graph Convolutional Networks (R-GCN) to learn intricate patent attribute relationships.
• Utilized Bidirectional Long Short-Term Memory networks (BiLSTM) to aggregate temporal evolution differences.
• Applied Principal Component Analysis (PCA) to analyze citation evolution characteristics.

Main Results:

• The PTNS model demonstrated superior performance compared to baseline methods.
• Significant reductions in Root Mean Squared Logarithmic Error (RMSLE) were observed for new, grown, and random patents (approx. 0.04, 0.14, 0.18).
• Mean Absolute Logarithmic Error (MALE) also decreased substantially for new, grown, and random patents (approx. 0.04, 0.12, 0.16).

Conclusions:

• The PTNS model effectively predicts patent citation trajectories by integrating temporal network analysis and deep learning.
• The model's ability to capture temporal dynamics and complex relationships enhances the identification of high-value patents.
• This approach offers a robust method for navigating the uncertainties in patent citation behavior.