High-Value Patents Recognition with Random Forest and Enhanced Fire Hawk Optimization Algorithm

High-value patents are a key indicator of new product development, the emergence of innovative technology, and a source of innovation incentives. Multiple studies have shown that patent value exhibits a significantly skewed distribution, with only about 10% of patents having high value. Identifying high-value patents from a large volume of patent data in advance has become a crucial problem that needs to be addressed urgently. However, current machine learning methods often rely on manual hyperparameter tuning, which is time-consuming and prone to suboptimal results. Existing optimization algorithms also suffer from slow convergence and local optima issues, limiting their effectiveness on complex patent datasets. In this paper, machine learning and intelligent optimization algorithms are combined to process and analyze the patent data. The Fire Hawk Optimization Algorithm (FHO) is a novel intelligence algorithm suggested in recent years, inspired by the process in nature where Fire Hawks capture prey by setting fires. This paper firstly proposes the Enhanced Fire Hawk Optimizer (EFHO), which combines four strategies, namely adaptive tent chaotic mapping, hunting prey, adding the inertial weight, and enhanced flee strategy to address the weakness of FHO development. Benchmark tests demonstrate EFHO's superior convergence speed, accuracy, and robustness across standard optimization benchmarks. As a representative real-world application, EFHO is employed to optimize Random Forest hyperparameters for high-value patent recognition. While other intelligent optimizers could be applied, EFHO effectively overcomes common issues like slow convergence and local optima trapping. Compared to other classification methods, the EFHO-optimized Random Forest achieves superior accuracy and classification stability. This study fills a research gap in effective hyperparameter tuning for patent recognition and demonstrates EFHO's practical value on real-world patent datasets.