Text

theses

Patent and patent citation networks are rich datasets. In this dissertation we develop graph mining algorithms for the analysis of patent citation networks. First we develop a measure of patent influence within a patent citation network. Identifying influential or important patents helps in decision making, including focusing investment. We propose algorithms based on the powerful graph kernels for the ranking of patents in influence, and we demonstrate how the von Neumann graph kernel is well suited for influence analysis in patent citation networks. Secondly, we present new similarity measures between patents in a patent citation network. In the past, techniques such as text mining and keyword analysis have been applied for patent similarity calculation. The drawback of these approaches is that they depend on word choice and writing styles of authors. In this work we develop new similarity measures for patents in a patent citation network using only the patent citation network structure. The proposed similarity measures use multi-stage co-citation and bibliographic coupling links. Applications of the similarity measures include outlier scoring of patents in patent citation networks. Finally, we propose new methods for scoring and ranking patents in outlierness within a patent citation dataset. A distinguishing characteristic of patent datasets is that they contain both attribute data describing patents, as well as graph structure data in the citation network. Traditional outlier ranking techniques usually focus on either homogeneous vector data or on graph structure data. In this work we propose new outlier ranking methods developed specifically for patents in an attributed patent citation network. One challenge is how outlier ranking should handle these two different data types in an integrated fashion. To address this challenge, we first develop a new patent subspace clustering algorithm that considers both types of data. Based on the patent clustering result, we then develop methods for the scoring and ranking of patents in outlierness within patent citation networks. Proposed outlier score functions consider both patent attribute data and graph structure data. We compare the performance of our developed approaches with existing approaches using synthetic data and real-life U.S. patent data.

ETD_6228

Ph.D.

Includes bibliographical references

by Andrew David Rodriguez

doi:10.7282/T3H133VP

ETD doctoral

The author owns the copyright to this work.