Unsupervised learning is a way to extract knowledge from noisy and complex sets of unlabeled data. The video-surveillance setting provides a potentially huge amount of unlabeled information on a given scene. In this paper we explore the use of spectral clustering to learn common behaviours from sets of dynamic events from a video-surveillance system. In particular we discuss how temporal data, characterized by variable lengths and an internal ordering, may be exploited effectively by means of appropriate representations and kernel functions. An experimental assessment on synthetic and real data guides us to an effective solution based on the use of strings.