Learning a succinct set of substructures that predicts global network properties plays a key role in understanding complex network data. Existing approaches address this problem by sampling the exponential space of all possible subnetworks to find ones of high prediction accuracy. In this paper, we develop a novel framework that avoids sampling by formulating the problem of predictive subnetwork learning as node selection, subject to network-constrained regularization. Our framework involves two steps: (i) subspace learning, and (ii) predictive substructures discovery with network regularization. The framework is developed based upon two mathematically sound techniques of spectral graph learning and gradient descent optimization, and we show that their solutions converge to a global optimum solution -- a desired property that cannot be guaranteed by sampling approaches. Through experimental analysis on a number of real world datasets, we demonstrate the performance of our framework against state-of-the-art algorithms, not only based on prediction accuracy but also in terms of domain relevance of the discovered substructures.