Automated customer management with sequential pattern mining of journey data logs

Abstract

The research proposes a framework for developing customer service business models. This process involves automated actions such as real-time customer sentiment analysis, response generation, and intelligent routing of customer queries based on analysing customer-business interaction logs. One of the key practical benefits of our proposed framework is service automation, which can significantly enhance decision-making. This can lead to the delivery of high-quality services and improved end-user satisfaction, and reduced business costs.

The primary research problem addressed is the early identification of customers who are likely to experience a significant increase in total customer journey time. This is a critical step towards resolving their issues by escalating to a human service interaction. This process occurs in the context of complex data, where customer journey models must be adapted over time to account for drifts in the statistical properties and behavioural patterns of customer journeys.

The central hypothesis of this thesis is that an adaptive system can more efficiently detect customers at risk of lengthy customer journeys. Based on event log data, an autonomic framework has been designed and tested to account for the concept drift of customer models over time. The framework considers competing customer journey segmentation approaches and performs periodic reclustering. Prediction models were developed based on linear and non-linear techniques via sequential pattern analysis and process mining event logs. Various clustering approaches were used and evaluated to model customer journeys.

Customer journey subsequences of varying lengths were investigated statistically. Adaptive customer segmentation and profiling were then used to identify and prioritise at-risk customer groups and resolve their problems at the initial contacts. This approach outperforms previous techniques, including Classification and Regression Trees, and Logistic Regression algorithms. As in those approaches, minimal hyperparameter tuning was required, and they were deterministic. However, proposed approaches are adaptive over time regardless of changes in business policies.

Thesis is embargoed until 31st January 2028

Date of Award	Jan 2026
Original language	English
Supervisor	Darryl Charles (Supervisor), Bryan Scotney (Supervisor) & Glenn Hawe (Supervisor)