Explainability in Process Outcome Prediction: Guidelines to Obtain Interpretable and Faithful Models

Both in operations research (OR) and business process management (BPM), prevalent topics include the modelling of processes in order to identify possible problems such as bottlenecks caused by a mismanagement or lack of resources [1] with the goal to find root causes in the process flow [2]. Over the past two decades, the BPM domain has seen a strong uptake of data-driven process analysis, coined under the term process mining, which uses process data generated by executed processes for cases within an information system [3]. This follows a similar trend in OR, where research shifted because of the access to large databases on (operational) transactions and a lack of back testing [4]. The focus of this study lies with predictive process monitoring [5], the umbrella term geared towards process mining for predictive activities. It allows identifying process-related trends regarding particular outcomes (e.g., will customers be awarded credit?), impeding bottlenecks (e.g., how long will it take to process my credit application?), and whether particular activities will occur in the future (e.g., will a credit check be necessary for this application?). When the concrete objective is to predict the outcome of an incoming, incomplete case, the field of study is referred to as Process Outcome Prediction (POP). The process data used in this research field is also referred to as event logs, as the occurrence of a single activity in a process (case) is referred to as'event'. Moreover, an event log consists of traces, each a sequence of events produced in the context of one case.