Improved Classi er Training Methods for Predictive Process Monitoring

Abstract

Recently, there has been numerous studies on the use of machine learning (ML) methods for business enhancement across different areas. Organiza- tions are in need to improve their business process performance by utilizing predictive models for monitoring ongoing business cases. Predictive process monitoring (PPM) tackles this problem by forecasting the behaviour, ex- ecution, and outcome of business processes at runtime. PPM approaches take an event log (i.e. a collection of completed cases) as input and utilize ML methods to train models to predict the future state of a given case, and to answer questions such as: Will a loan application will be approved or declined (i.e. nal outcome)? What is the next event given the previous events? Or what is the remaining time until the end of the case? A speci c, family of approaches of PPM, known as outcome-oriented PPM, focuses on predicting whether or not a case will end with an expected outcome or not. An outcome-oriented PPM framework is expected to form precise predic- tions in the early execution stages to decide if the system worker should take part and get involved or not and to avoid unexpected outcomes. In this setting, this thesis addresses the question of how to improve the pre- dictive process monitoring of business process outcomes. To answer this question, we propose three different enhancements to the currently existing approaches that have been introduced in the literature. The proposed en- hancements are evaluated using a benchmark covering 20 prediction tasks that come from different real-life event logs. Empirical results con rm that our proposed approaches deliver signi cant improvements relative to exist- ing PPM techniques in terms of accuracy.