Clustering algorithms performance analysis applied to patent database

Abstract

The granularity of large patent classification systems hampers the reclassification process in which patent categories are broken down into smaller ones, suggesting new categories. As these groups belong to a constricted domain of knowledge, keywords and subject descriptors tend to be similar and therefore insufficient to differentiate documents. In this context, the identification of common cited references can be useful to define semantic relationship among patents. This work compares citation analysis based results obtained by three clustering algorithms, SOM networks, K-Means and Multi-SOM. An empirical experiment was conducted using a patent database from the United States Patent and Trademark Office with all patents of four subgroups classified by the Cooperative Patent Classification system. Practical results evaluated by statistical inference techniques showed that SOM performs better than the other algorithms to cluster that database. This study can contribute with the reclassification process for a subgroup level of current patent classification systems, demonstrating how citation analysis can be an alternative attribute to the automatic clustering process.