Monday, March 12, 2012

CHAN Tian Heong, MIHM Jürgen, SOSA Manuel
A Structured Approach to Identifying Styles in Designs
INSEAD Working Paper 2012/29/TOM

A style is a set of designs that are visually similar to each other, yet distinctive from designs in other style sets. This paper presents a structured approach to identifying styles using cluster analysis. The basic premise of our approach is to build a similarity relationship between designs based on the idea that new designs ”cite” previous ones; they can be thought of as a recombination of past designs. Designs that draw from a common pool of previous designs are likely to be very similar. A style is formed by a set of highly similar designs. We apply a graph clustering algorithm to a dataset of 365,444 US design patents (from year 1976 to 2009) to group designs into styles. Design patents provide a good platform for our approach, in that both the visual aspects of the design as well as citation information are captured in the patent documentation. We test the output of our clusteringbased approach with two experiments using human subjects. We find that the outputs from the algorithm and the human subjects are in agreement in the sense of mismatch (i.e., if forced to group elements into a given set of clusters, humans and the algorithm make the same choices). We also found empirical support for the methods of how the clustering algorithm decides on the number of clusters. Implications for research on design theory are discussed.