It is difficult to relate the structure of a cortical neural network to its dynamic activity analytically. Therefore we employ machine learning and data mining algorithms to learn these relations from sample random recurrent cortical networks and corresponding simulations. Inspired by the PageRank and the Hubs & Authorities algorithms, we introduce the NeuronRank algorithm, which assigns a source value and a sink value to each neuron in the network. We show its usage to extract structural features from a network for the successful prediction of its activity dynamics. Our results show that NeuronRank features can successfully predict average firing rates in the network, and the firing rate of output neurons reflecting the network population activity. (c) 2006 Elsevier B.V. All rights reserved.