How to identify high-value innovations? A perspective from construction engineering patents' knowledge graphs

Purpose – To enhance the early prediction of high-value innovations in the construction engineering domain by addressing the limitations of existing approaches that overlook semantic and relational features embedded in patent data. Design/methodology/approach – This paper constructs a multidimensional predictive indicator system by integrating structural features with network relationships into a unified semantic framework of a patent knowledge graph in the construction engineering domain. An improved TF-IDF algorithm is applied to extract key technologies, while knowledge graph embedding techniques are used to capture latent relational features. Finally, machine learning classifiers are employed to predict patent value. Findings – The results indicate that combining patent text features with knowledge graph embedding features significantly enhances prediction performance. The model achieves an AUC exceeding 80%, representing an improvement of approximately 7% compared with models relying solely on external features. Additionally, the patent technology coverage and novelty features extracted from the patent text play a significant role in value prediction. Originality/value – This research integrates semantic, structural, and relational features within a knowledge graph-based approach for the early identification of high-value innovations. It highlights the theoretical role of relational networks and offers interpretive insights to inform innovation evaluation, serving as an empirical reference for R&D and policy analysis.