A framework for the analysis and comparison of process mining algorithms

Process mining algorithms use event logs to learn and reason about business processes. Although process mining is essentially a machine learning task, little work has been done on systematically analysing algorithms to understand their fundamental properties, such as how much data is needed for confidence in mining. Nor does any rigorous basis exist on which to choose between algorithms and representations, or compare results. We propose a framework for analysing process mining algorithms.

Processes are viewed as distributions over traces of activities and mining algorithms as learning these distributions. We use probabilistic automata as a unifying representation to which other representation languages can be converted.

To validate the theory we present analyses of the Alpha and Heuristics Miner algorithms under the framework, and two practical applications. We propose a model of noise in process mining and extend the framework to mining from ‘noisy’ event logs. From the probabilities and sub-structures in a model, bounds can be given for the amount of data needed for mining. We also consider mining in non-stationary environments, and a method for recovery of the sequence of changed models over time.

We conclude by critically evaluating this framework and suggesting directions for future research.