Herlau, Tue3; Mørup, Morten3; Schmidt, Mikkel Nørgaard3; Hansen, Lars Kai4
1 Department of Informatics and Mathematical Modeling, Technical University of Denmark2 Cognitive Systems, Department of Informatics and Mathematical Modeling, Technical University of Denmark3 Department of Applied Mathematics and Computer Science, Technical University of Denmark4 Copenhagen Center for Health Technology, Center, Technical University of Denmark
Relational modelling classically consider sparse and discrete data. Measures of influence computed pairwise between temporal sources naturally give rise to dense continuous-valued matrices, for instance p-values from Granger causality. Due to asymmetry or lack of positive definiteness they are not naturally suited for kernel K-means. We propose a generative Bayesian model for dense matrices which generalize kernel K-means to consider off-diagonal interactions in matrices of interactions, and demonstrate its ability to detect structure on both artificial data and two real data sets.