A0235
Title: Asymmetric cluster difference scaling based on hill-climbing model
Authors: Kensuke Tanioka - Doshisha University (Japan) [presenting]
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: Asymmetric (dis)similarity data is dissimilarity data such that dissimilarity from subject $i$ to subject $j$ does not necessarily match the dissimilarity from subject $j$ to $i$. In fact, asymmetric (dis)similarity data is observed in various situations such as brand switching, or network analysis for SNS. Given asymmetric (dis)similarity data, Asymmetric Multidimensional Scaling (AMDS) is a very useful tool to interpret the asymmetric relations between subjects visually. AMDS provides us with two things on the lower dimension; One is the coordinates of each subject, and the other is parameters for describing asymmetric relations. There are various AMDS have been proposed, and these methods depend on the model of parameters for the asymmetric relation. However, these days, information technology is improved, and we have to deal with large and complex data. If AMDS is applied to such large asymmetric (dis)similarity data, it becomes difficult to interpret the asymmetric relation because the number of subjects is large. To overcome the problem, we propose new AMDS for the cluster centroids, not subjects. The proposed AMDS visualizes asymmetric relations between clusters. In the proposed method, the hill-climbing model is adopted from the candidates of AMDS models.
