Abstract:
Process mining consists of extracting knowledge and actionable information
from event-logs recorded by Process Aware Information Systems (PAIS).
PAIS are vulnerable to system failures, malfunctions, fraudulent and undesirable executions resulting in anomalous trails and traces. The flexibility
in PAIS resulting in large number of trace variants and the large volume of
event-logs makes it challenging to identify anomalous executions and deter-
mining their root causes. We propose a framework and a multi-step process
to identify root causes of anomalous traces in business process logs. We fi rst transform the event-log into a sequential dataset and apply Window-
based and Markovian techniques to identify anomalies. We then integrate
the basic eventlog data consisting of the Case ID, time-stamp and activity
with the contextual data and prepare a dataset consisting of two classes
(anomalous and normal). We apply Machine Learning techniques such as
decision tree classifi ers to extract rules (explaining the root causes) describing anomalous transactions. We use advanced visualization techniques such
as parallel plots to present the data in a format making it easy for a process analyst to identify the characteristics of anomalous executions. We
conduct a triangulation study to gather multiple evidences to validate the
effectiveness and accuracy of our approach.
