Machine Learning and Sentiment Analysis with application to Customer Satisfaction

In the constantly changing landscape of business, customer satisfaction stands as a cornerstone of success. To interpret the intricate nuances of consumer feedback and sentiments, organizations frequently turn to Likert-type scales, a popular tool for collecting structured data. However, the inherent subjectivity of Likert-type responses introduces uncertainty and imprecision within such data. Beside such structured data, unstructured textual data emerged as a powerful source of customer satisfaction insights. Leveraging natural language processing techniques, organizations can mine unstructured textual feedback to derive customer sentiments and opinions. This qualitative dimension complements quantitative ratings, offering a holistic view of consumer satisfaction. This thesis explores innovative approaches to augment customer satisfaction analysis by addressing the challenges of comprising uncertainty within rating data and of combining textual and rating data. The propose methodologies lie in the context of clustering techniques highlighting how clustering enables the identification of consumer segments with similar preferences and experiences, fostering tailored strategies and resource allocation. After a comprehensive and structured literature review some research questions are identified and the thesis try to solve such research problem by developing paths of analysis that adopt and combine different statistical and machine learning techniques: Combination of discrete Uniform and shifted Binomial random variable model (CUB model), Fuzzy sets theory, cluster analysis, sentiment and emotion analysis and topic modeling. In the main results an Adaptive membership function (MF) for data fuzzification is introduced exploiting CUB model uncertainty to handle the subjectivity and uncertainty entailed within responses measured on Likert-type scales. The Adaptive MF approach mitigates the subjectivity in the experts' definition of the membership functions, allowing the incorporation of respondent-specific uncertainty into the analysis. A second method develops a clustering algorithm for mixed data called FCMd-CUB that allow to partition consumer based on the rating they express and the feeling and uncertainty derived from a CUB model. Lastly a process of analysis combining textual data and rating data (with particular focus on review web data) is presented. All proposed methodology of analysis are supported by application on real case study data and/or simulation study. The thesis concludes by emphasizing the practical utility of these methodologies in enhancing customer satisfaction analysis. By embracing advanced techniques that combine structured and unstructured data, organizations can gain a deeper understanding of consumer sentiments, improve decision-making, and tailor strategies to individual preferences.