Hybrid machine learning and MCDM framework for consumer preference extraction and decision support in dynamic markets

Abstract

With the rapid development of e-commerce and digital consumption, online reviews have become an important channel for consumers to express their opinions and for businesses to understand market dynamics. However, the surge in review volume—resulting in information overload and a large amount of irrelevant content—has severely hindered the accurate identification of genuine consumer preferences, directly affecting the accuracy of product design, marketing, and resource allocation decisions. To address this challenge, this paper proposes an innovative hybrid framework that integrates information entropy, a binary classification model, PCA combined with K-means clustering, the BERT-wwm-ext sentiment analysis model, and multi-criteria decision-making (MCDM) methods, aiming to enhance the accuracy of preference analysis and the reliability of decision-making in the context of digital consumption. The framework tackles three key challenges: bias in traditional evaluations of perceived useful information, incompleteness in preference feature extraction, and inaccuracies in preference weight calculation. A comprehensive analysis of over 70,000 online customer reviews sourced from platforms such as the Apple App Store and JD.com validates the framework, showing that it outperforms existing models in predicting perceived usefulness, uncovering hidden product attributes, and refining feature weight calculations. This study not only provides robust data support for enterprises in product optimization and targeted marketing, but also offers decision makers a scientifically grounded framework for product management and efficient resource allocation that better aligns with consumer needs.