Collective classification refers to the classification of interlinked and relational objects described as nodes in a graph. The Iterative Classification Algorithm (ICA) is a simple, efficient and widely used method to solve this problem. It is representative of a family of methods for which inference proceeds as an iterative process: at each step, nodes of the graph are classified according to the current predicted labels of their neighbors. We show that learning in this class of models suffers from a training bias. We propose a new family of methods, called Simulated ICA, which helps reducing this training bias by simulating inference during learning. Several variants of the method are introduced. They are both simple, efficient and scale well. Experiments performed on a series of 7 datasets show that the proposed methods outperform representative state-of-the-art algorithms while keeping a low complexity.