Strategic Citations in Patents: Analysis Using Machine Learning

Abstract

As an alternative to patent citations, I construct a measure of proximity in ideas expressed in patent text using unsupervised machine learning algorithm Doc2Vec. Patent citations are the most commonly used indicator of knowledge relationships across patents. However, citation may be prone to strategic behaviour: inventors and firms may over-cite to offset the likelihood of litigation, and under-cite to increase the scope of the patent. Based on the similarity across patent texts, I find evidence that (i) applicants may strategically omit citations to patents in different cities, as infringement discovery is less likely; (ii) applicants cite their own prior inventions less after they change firms. This implies that use of citations to measure knowledge flows are affected by strategic biases. The novel approach of this paper is to use machine learning methods to categorise and analyse similarity for over one million patents in order to uncover behavioural biases in inventors' citation patterns.