A novel comprehensive method for customer segmentation based on identifying topics and sentiments from unstructured online product reviews

In product development and business insights, topic extraction and sentiment analysis are crucial components. Due to the information overload in e-commerce reviews and the diverse preferences of customers, traditional research methods fail to identify commonalities among customers effectively. To overcome these challenges, we proposed an innovative five-stage ensemble approach for customer segmentation. First, TextRank was employed for data preprocessing to extract key textual features and filter relevant content. Subsequently, key topics were identified through the Word2Vec-based topic identification model. Then, to enhance the accuracy of topic-level sentiment scores, clause-level sentiment analysis was conducted using BERT, where sentiment scores were fine-tuned through TF-IDF weighting for enhanced granularity. After that, interpretable machine learning (IML) algorithms were employed to analyze user satisfaction (USAT), ensuring predictive performance and model transparency. Finally, deep embedded clustering (DEC) was leveraged to perform customer segmentation based on the extracted key topic-sentiment features. The effectiveness of the proposed method was validated through a real-world case study involving 22, 320 online user reviews. The results showed that categorical boosting (CatBoost) achieved the highest performance, with an F1-score of 0.9433, demonstrating its high accuracy and transparency in predicting USAT determinants. The findings facilitate the identification of innovative product concepts.