Analysing and Modelling Employee Turnover using Causal Bayesian Networks: Mixed Evidence Approach for Causal Knowledge Discovery

Abstract The purpose of this thesis is to explore the practicalities of using causality to build a decision model for Employee Turnover. In previous research, Employee Turnover was analysed using
supervised machine learning to improve its prediction aspect, however, having a decision model for turnover might be of use for human resources practitioners more than prediction models to support their decisions.
This research therefore investigated Causal Bayesian Networks (CBNs) as a modelling technique to analyse Turnover. CBNs allow the incorporation of domain experts knowledge and data-driven insights. They are probabilistic graphical models that capture dependencies among variables, their defining features include managing uncertainty through probability theory and using directed edges to represent cause-and-effect relationships.
The major contribution of this thesis is analysing the practicalities of applying causality/ causal discovery to build a decision model for Employee Turnover. This is the first research to consider modelling Employee Turnover from a causal discovery perspective (using CBNs) based on mixed evidence. The main focus of this research was to incorporate flexibility into scenario planning, allowing HR managers to test multiple scenarios. The key contributions of this study were as follows: a new hybrid approach was introduced to capture core causal knowledge from domain experts for use with structure learning algorithms, an evidence comparison and a qualitative review was conducted using insights from both academic literature and semi-structured interviews with senior HR managers, a comparison between different classes of structure learning algorithms to generate Directed Acyclic Graphs (DAGs) from HR data, three DAGs were developed using domain experts knowledge and structure learning algorithms and three CBNs were evaluated for scenario planning about turnover.
These contributions collectively supported the development of a framework for building CBNs in a transparent way, leveraging expert knowledge and structure learning algorithms.