Thesis

Developing an improved complaint management system through the use of text modelling techniques

Conteúdo disponível para baixar

Baixar PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2025
Thesis identifier
  • T17421
Person Identifier (Local)
  • 201771817
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • This thesis investigates the integration of textual complaint data into customer churn prediction (CCP) models, addressing a critical gap in the intersection of service recovery and predictive analytics. While traditional CCP models primarily rely on structured variables, textual complaints encapsulate rich customer sentiments and behavioural signals that remain underutilised in churn modelling. This research, therefore, explores how advanced text representation techniques can enhance CCP models, bridging the gap between service failure recovery theories and data-driven predictive modelling. Grounded in service recovery and customer relationship management (CRM) theories, this study evaluates a decision-support framework that incorporates textual complaints and structured variables to improve churn prediction. Using real-life customer complaint data from a UK-based data-driven product company, the research benchmarks traditional count-based text representations (e.g., TF-IDF) against modern embedding-based methods (e.g., word embeddings, Transformer models). Additionally, the study investigates data fusion techniques, examining their role in leveraging multimodal information for improved churn prediction. Key findings reveal that: 1. Incorporating textual complaint data significantly enhances CCP models, confirming the value of textual analysis in understanding customer behaviours. 2. Word embedding models outperform TF IDF-based models in overall CCP performance, indicating a shift towards more sophisticated text representation techniques. 3. TF-IDF-based models perform better at predicting retained customers, while word embedding models excel in identifying churn instances, underscoring the importance of task-dependent model selection. 4. An ensemble approach combining count-based and latent feature representations improves retained case prediction but slightly decreases churn prediction accuracy, suggesting that model selection should align with specific business objectives. 5. Data fusion techniques play a crucial role in optimizing predictive accuracy, demonstrating the need for well-designed multimodal integration strategies. 6. Structured variables remain essential in CCP models, providing additional insights into customer retention dynamics. This thesis advances both theoretical and practical understanding of customer churn prediction by demonstrating how textual complaint data can be strategically leveraged to enhance CCP models. It contributes to service recovery and predictive analytics literature by offering a data-driven approach to customer retention strategies, equipping businesses with more effective and intelligent service recovery mechanisms.
Advisor / supervisor
  • Quigley, John
  • Wilson, Alan
Resource Type
DOI
Date Created
  • 2023

Relações

Itens

Miniatura Título Data de carga Acesso Ações
file details PDF of thesis T17421 2025-07-08 Público Baixar