Abstract
We apply the network Lasso to solve binary classification and clusteringproblems for network-structured data. To this end, we generalize ordinarylogistic regression to non-Euclidean data with an intrinsic network structure.The resulting "logistic network Lasso" amounts to solving a non-smooth convexregularized empirical risk minimization. The risk is measured using thelogistic loss incurred over a small set of labeled nodes. For theregularization, we propose to use the total variation of the classifierrequiring it to conform to the underlying network structure. A scalableimplementation of the learning method is obtained using an inexact variant ofthe alternating direction methods of multipliers which results in a scalablelearning algorithm