A method and system are described for end-user transaction recognition based on server data such as sequences of remote procedure calls (RPCs). The method may comprise machine-learning techniques for pattern recognition such as Bayesian classification, feature extraction mechanisms, and a dynamic-programming approach to segmentation of RPC sequences. The method preferably combines information-theoretic and machine-learning approaches. The system preferably includes a learning engine and an operation engine. A learning engine may comprise a data preparation subsystem (feature extraction) and a Bayes Net learning subsystem (model construction). The operation engine may comprise transaction segmentation and transaction classification subsystems.