Context-Aware Prediction on Business Process Executions Francesco Folino, Massimo Guarascio, and Luigi Pontieri Institute for High Performance Computing and Networking (ICAR) National Research Council of Italy (CNR) Via Pietro Bucci 41C, I87036 Rende (CS), Italy {ffolino,guarascio,pontieri}@icar.cnr.it Abstract. Discovering predictive performance models is an emerging topic in Process Mining. However, making accurate estimates is not easy especially when considering fine-grain metrics (such as processing times) on complex and flexible processes, where performances may change over time depending on context factors. We try to face such a situation by a general predictive-clustering approach, where different context-related execution scenarios are find and equipped with distinct performance- prediction models. A two-stage forecast can be then made for a new process case by using the model of the cluster it is estimated to belong to. Tests on real-life logs confirmed the validity of the approach. 1 Introduction Process mining techniques [6] are a precious tool for analysing business processes, owing to their capability to extract useful information out of historical process logs. While most traditional approaches focused on the discovery of control- flow models, increasing attention has been gained by the discovery of predictive process models, providing operational support at run-time. In particular, an emerging research stream [7, 5] concerns the induction of models for forecasting performances metrics on new process instances. However, accurate forecasts are not easy to make for fine-grain measures (like, e.g., processing times), especially when the analyzed process exhibits complex and flexible dynamics, and its execution schemes and performances change over time, depending on the context. In general, more precise process models can be found by exploiting ad-hoc clustering methods [4], while regarding each resulting cluster as evidence for a peculiar execution scenario. However, all previous clustering-oriented process mining approaches focused on control-flow aspects, without spending any effort towards improve performance predictors. In this paper, which summarizes and generalizes the approach in [3], we de- scribe a general predictive-clustering computation scheme, meant to detect differ- ent context-related execution scenarios (or process variants ), and to equip each of them with a process performance model. The final outcome is a novel kind of