Inventor Name Disambiguation with Gradient Boosting Decision Tree and Inventor Mobility in China (1985-2016)

This paper presents the first systematic disambiguation result of all Chinese patent inventors in the State Intellectual Property Office of China (SIPO) patent database from 1985 to 2016. We provide a method of constructing high-qualitative training data from lists of rare names and evidence for the reliability of these generated labels when large-scale and representative hand-labeled data are crucial but expensive, prone to error, and even impossible to obtain. We then compare the performances of seven supervised models, i.e., naive Bayes, logistic, linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), as well as tree-based methods (random forest, AdaBoost, and gradient boosting decision trees), and found that gradient boosting classifier outperforms all other classifiers with the highest F1-score and stable performance in solving the homonym problem prevailing in Chinese names. In the last step, instead of adopting the more popular hierarchical clustering method, we clustered records with the density-based spatial clustering of applications with noise (DBSCAN) based on the distance matrix predicated by the GBDT classifier. Varying across different testing data and parameters of DBSCAN, our algorithm yielded a F1-score ranging from 93.5%-99.3% with splitting error within the range 0.5%-3% and lumping error between 0.056%-0.37%. Based on our disambiguated result, we provide an overview of Chinese inventors' regional mobility.