Combining Natural Language Processing and Federated Learning for Consumer Complaint Analysis: A Case Study on Laptops

Customer feedback is an invaluable source of information for any organization, and crucial business phases such as market research, product development, and post-sale services may greatly benefit from it. Existing methods utilize qualitative data, surveys, and excel-generated reports to analyze customer complaints and feedback. However, it takes a while to gather information, classify appropriately, and analyze it manually before the data becomes useful information for the organization and the complaints are resolved simultaneously. To address these challenges, NLP is applied in this paper to categorize and examine user reviews and feedback using a laptop user feedback data set. The federated learning technique is used to ensure the privacy of the participating clients. In addition, it can properly handle uneven data distribution and lessen the load on the central server. After each federated training round, the FedAvg algorithm has been used to incorporate the local model learning into the central model. Three BERT variants (BERT, DistilBERT, and RoBERTa) have experimented with federated learning for the classification task. The best result was obtained by the RoBERTa variant (71.59% accuracy for the independent and identical data distribution setting, 71.55% for the non-independent and non-identical setting). Therefore, this paper proposes a comprehensive system to produce product summaries and customer discontent reports based on graphical representations of the product and NLP-assisted classification of user reviews, complaints, and feedback according to their usage history. In the web application, the best-performing model variant is deployed and used. The simulation of the web application shows the successful execution of the proposed system. © 2023, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.