EigenPatent: A novel eigenvector-based ranking method for enhanced patent similarity detection

This is the authors' abstract. We don't add key points for this paper.

Patent retrieval often faces unique challenges due to the complex structure and nature of technical documents. Traditional similarity measures often fail to capture the nuanced semantic relationships between inventions, marking it difficult for non-experts to retrieve relevant prior art. This study introduces a novel eigenvector-based ranking methodology for patent similarity that significantly outperforms traditional embedding approaches. We integrate covariance-matrix analysis with hyperplane projections, to capture both semantic and structural relationships between technical documents. Experiments within the F03D patent subclass demonstrate our approach achieves a similarity score of 83.37, substantially outperforming Word2Vec (23.3), ELMo (21.1), and SimCSE (44.8). This work addresses critical challenges in patent retrieval while introducing innovations applicable to broader technical document similarity tasks, enabling non-experts to efficiently identify relevant prior work without specialized knowledge of patent systems. • Novel eigenvector-based ranking methodology for patent similarity outperforms traditional embedding approaches. • Integration of covariance-matrix analysis with hyperplane projections captures semantic and structural relationships. • Achieves 83.37 similarity score in F03D patent subclass, substantially outperforming Word2Vec, ELMo, and SimCSE. • Introduces domain-specific evaluation metrics for patent similarity with comprehensive benchmarking protocols. • Provides scalable solution balancing computational efficiency and ranking performance for technical document retrieval.