K.U.Leuven - Faculty of Economics and Applied Economics
DTEW - KBI_0716 pages:1-20
Process mining is the automated acquisition of process models from the event logs of information systems. Although process mining has many useful applications, not all inherent difficulties have been sufficiently solved. A first difficulty is that process mining is often limited to a setting of non-supervised learnings since negative information is often not available. Moreover, state transitions in processes are often dependent on the traversed path, which limits the appropriateness of search techniques based on local information in the event log. Another difficulty is that case data and resource properties that can also influence state transitions are time-varying properties, such that they cannot be considered ascross-sectional.This article investigates the use of first-order, ILP classification learners for process mining and describes techniques for dealing with each of the above mentioned difficulties. To make process mining a supervised learning task, we propose to include negative events in the event log. When event logs contain no negative information, a technique is described to add artificial negative examples to a process log. To capture history-dependent behavior the article proposes to take advantage of the multi-relational nature of ILP classification learners. Multi-relational process mining allows to search for patterns among multiple event rows in the event log, effectively basing its search on global information. To deal with time-varying case data and resource properties, a closed-world version of the Event Calculus has to be added as background knowledge, transforming the event log effectively in a temporal database. First experiments on synthetic event logs show that first-order classification learners are capable of predicting the behavior with high accuracy, even under conditions of noise.