Title: Bayesian sparse convex clustering via NEG distribution
Authors: Kaito Shimamura - NTT Advanced Technology Corporation / The University of Electro-Communications (Japan) [presenting]
Shuichi Kawano - The University of Electro-Communications (Japan)
Abstract: Sparse convex clustering is convex clustering, which is a convex relaxation of classical clustering methods, with variable selection. Sparse regularization plays a key role of selecting relevant variables in sparse convex clustering. In sparse convex clustering, we need to set values of weights in the regularization term. By setting the values properly, it is known that the accuracy of clustering and variable selection improves. However, it is pointed out that the values highly depend on observed data. This causes a degradation of estimation accuracy in sparse convex clustering when the sample size is small. To overcome the problem, we first introduce a Bayesian formulation of sparse convex clustering. We then propose a Bayesian sparse convex clustering based on a normal-exponential-gamma (NEG) prior distribution. We conduct numerical studies to examine the effectiveness of the Bayesian model.