Actionable recommendations to enhance the win-back process model using machine learning and data analysis applied on customer history, demographics and behavior during process

 

Actionable recommendations to enhance the win-back process model using machine learning and data analysis applied on customer history, demographics and behavior during process

Student: Yasmin ElSawy

Title: Actionable recommendations to enhance the win-back process model using machine learning and data analysis applied on customer history, demographics and behavior during process

Supervisor: Alessandro Berti

1st Examiner: Prof. Wil M.P. van der Aalst

2nd Examiner: Prof. Dr. Ulrik Schroeder

Summary

Over the last couple of years, declarative models are used to present the process by capturing the high-level behavior in the process. Consequently, an easy interpretation of the process behavior is provided to the stakeholders. Current research mainly focuses on discovering constraints from the event log which explains the process in action. While the thesis proposed approach focuses on discovering constraints that increase the chances to achieve the process goal provided binary labeled traces. The approach proposed combines topics from machine learning and process mining. The constraints generated suggest an enhancement to the process to allow for more positive class labels. The machine learning models used in the proposed approach are used to explain how each constraint affects the process goal. Two approaches were followed to discover constraints provided an event log. The first approach is using a decision tree, where each node represents a constraint and the leaves represent the class labels. The second approach is using the XGBoost machine learning model which is interpreted using SHAP to transform the model black box into an interpretable model. The approach is evaluated on event logs of a Telecommunication company that contains the customer’s journey after submitting his contract cancellation request. The results show that the XGBoost combined with SHAP provides constraints that result in positive class labels more than negative class labels. The approach is also applied to the BPI Challenge event logs to show that it is generic and can also be applied to different event logs and domains as long as there is a binary target in the process and defined process goal.