Random Forest DBSCAN for USPTO Inventor Name Disambiguation

Abstract

Name disambiguation and the subsequent name conflation are essential for the correct processing of person name queries in a digital library or other database. It distinguishes each unique person from all other records in the database. We study inventor name disambiguation for a patent database using methods and features from earlier work on author name disambiguation and propose a feature set appropriate for a patent database. A random forest was selected for the pairwise linking classifier since they outperform Naive Bayes, Logistic Regression, Support Vector Machines (SVM), Conditional Inference Tree, and Decision Trees. Blocking size, very important for scaling, was selected based on experiments that determined feature importance and accuracy. The DBSCAN algorithm is used for clustering records, using a distance function derived from random forest classifier. For additional scalability clustering was parallelized. Tests on the USPTO patent database show that our method successfully disambiguated 12 million inventor mentions within 6.5 hours. Evaluation on datasets from USPTO PatentsView inventor name disambiguation competition shows our algorithm outperforms all algorithms in the competition.

Cited by in corpus (4)

• LAGOS-AND: A Large Gold Standard Dataset for Scholarly Author Name Disambiguation
• Optimal Transport-based Alignment of Learned Character Representations for String Similarity
• Estimating the Performance of Entity Resolution Algorithms: Lessons Learned Through PatentsView.org
• Representing text as abstract images enables image classifiers to also simultaneously classify text