Alpha Precision: Estimating the Significant System Behavior in a Process Model

One of the goals of process discovery is to construct, from a given event log, a process model which correctly represents the underlying system. As with any abstraction, one does not necessarily want to represent all possible behavior, but only the significant behavior. While various discovery algorithms support the use case of discovering the significant process behavior, proper evaluation measures for this use case appear to be missing.

In this talk, the results are presented of recent work on a new precision metric that quantifies to what extent the discovered model contains significant system behavior. Besides being a metric with a clear and intuitive interpretation, our new metric distinguishes itself in two other areas. Firstly, it introduces the concept of alpha-significance, which only measures precision with respect to significant behavior. Secondly, it is designed as a system measure and estimates the precision with respect to the underlying unobserved system rather than the observed log.

During the talk, the new precision measure and a statistical estimation method will be discussed. Additionally, an empirical demonstration and evaluation of the metric will be provided, which creates initial insights and knowledge about the performance and characteristics of the new measure.