To accomplish in their tasks, agents need to build models of other agents from observations. In open or adversarial settings, the observer agent does not know the full behavior repertoire of observed agents, and must learn a model of the other agents from its observations of their actions. This paper focuses on learning models of sequential behavior based on observed execution traces. It empirically compares sequence recognition approaches, and shows that they suffer from common deficiencies, including length-biases and inability to generalize discovered patterns. We present bias-removing and clustering methods to address these challenges, and evaluate them using synthetic and real-world data. The results show significant improvements in all the learning algorithms tested.