With the global rising Internet demand, network operators and service providers need to manage increasingly complex and interdependent systems. In this thesis, we explore how statistical learning techniques can be used to help modeling and understanding large computer networks. In a first contribution, we propose and evaluate a Graph Neural Network algorithm for path performance prediction given known network characteristics. Our second contribution focuses on Internet-scale Root Cause Analysis: given limited knowledge about the network, we evaluate three statistical learning techniques for this important problem, including Naive Bayes, Random Forest and Convolutional Neural Network classifiers. We show the results of these techniques over ...