Can’t see the forest for the leaves: Similarity and distance measures for hierarchical taxonomies with a patent classification example

Current measures of technological distance or similarity typically ignore a great deal of the information contained within the classification taxonomies upon which they are based. In this paper, I introduce two modifications which enable various common management research methods to fully make use of hierarchical classification data. Although the methods presented have broad applications, an extensive example exploring various measures of technological similarity based on the USPTO patent classification system is included. In addition to the general benefit of allowing the use of taxonomical data and thus more accurately reflecting the theoretical complexity of any underlying phenomenon, this methodology offers a number of other benefits specific to the patent context. These include the ability to model technological space within fields, since it correctly uses USPTO subclass level data, as well as the ability to accurately analyze similarity at the patent-to-patent dyadic level, since it calculates conceptual overlaps at the lowest level of the classification taxonomy. I explore the performance of these methods in two research contexts: (1) a patent-to-patent level sample within a single technological field and (2) an organization-to-organization level sample across industries. The results show that taxonomical methods generate more meaningful distributions of similarity scores within both samples and that similarity scores calculated via taxonomical methods have a more consistent relationship with citation likelihood and number of citations. Suggestions are provided for variants of this methodology for various technological and industrial classification systems and implications are drawn for future research in a wide range of domains.

► Challenges the validity of similarity measures that ignore hierarchical classification data.
► Offers easy transformations that allow any similarity/distance measure to appropriately utilize taxonomical classification data.
► Provides a method (as well as the classification dataset) to appropriately utilize USPTO subclass level classification data.
► Provides a method to accurately calculate technological distance between patents (i.e., at the patent-to-patent dyadic level of analysis).
► Shows that taxonomical measures of technological similarity have more meaningful relationship to citation likelihood in two samples.