Machine learning applied to Voice of Customer for feedback classification

This thesis presents a case study conducted during my internship at Toyota Motor Europe, with a primary focus on developing an algorithm to automatically classify negative customer feedback. The aim is to ensure that such feedback is appropriately directed to the relevant division and design group within the company, also assigning a Title and a short problem description. The algorithm was trained using machine learning techniques and evaluated on a large dataset of customer feedback collected from an external company, which specializes in providing VoC (Voice of Customer) data for the automotive sector. It is worth noting that the dataset used for training was labeled manually by employees over the past years, indicating a supervised learning approach. The model was subsequently adapted to textual data following a comprehensive preprocessing phase aimed at enhancing data quality and relevance. This preprocessing involved various steps, including data cleaning, normalization, and feature extraction, to ensure that the algorithm could effectively handle the complexities of natural language in customer feedback. The classifier's performance was assessed using a range of metrics, with particular emphasis on accuracy and precision. The results indicate that the proposed algorithm significantly decreases human workload while expanding the classification to include a broader range of models previously overlooked due to resource constraints. To ensure all employees, despite their technical knowledge, can easily use the algorithm, a user-friendly interface was developed using Flask, a web development microframework. Today, the proposed algorithm and user interface are utilized by Toyota Motor Europe's quality division to automate VoC analysis and facilitate data-driven decisions aimed at enhancing customer satisfaction. Effectively classifying Toyota Title and Problem categories, each with numerous classes, remains a notable challenge for future enhancement. Potential solutions may involve refining data preprocessing methodologies or integrating more sophisticated natural language processing techniques such as Large Language Models (LLMs).