Activity-Oriented Causal Process Mining: An End-to-End Approach Utilizing Ylearn

Abstract

In recent decades, companies have explored data-driven methods and tools to improve their business processes. More recently, prescriptive business process analysis became popular among data analysts and researchers. There are many studies on the use of prescriptive algorithms for the optimization of a variety of different business processes. Prescriptive algorithms given the historical and/or real-time data try to discover and recommend the best actions to improve the future outcome, e.g. what existing actions in the advertisement process need to be changed to increase the sales. One of the prescriptive methods approaches is Causal Process Mining which uses event logs received from the company’s information systems and then analyses them with Causal Inference algorithms to discover and estimate these possible changes (treatments) that would affect the final outcome. However, all event logs can differ by the variables that are logged and the models may become dependent on the data structure. This means that each event log requires separate variables investigation and modeling that would match the event log data structure. Consequently, performing these activities takes time and resources. A more generic and automated approach could be better applicable in different business cases and give useful results without excessive analysis or model building. For this reason, in this study, we investigate the possibility to use only case ID, activity, and timestamp variables of the event log for the causal inference algorithms. We propose the experimentation software artifact that includes data preparation and integrates the existing Ylearn causal inference tool. The approach is evaluated using five real-world event logs. Evaluation results show that causal relationships can be detected between activities of the event log and estimated treatment effects are comparable with other approaches.