Discovering the relationship between attributes of facial masks and review rating in online customer reviews using Explainable Artificial Intelligence (XAI)

Abstract

The COVID-19 pandemic led to a surge in facial mask usage. Analyzing customer feedback is crucial for enhancing this product, given the abundance of options and online reviews available. Techniques like Latent Dirichlet Allocation (LDA) combined with machine learning (ML) models help identify product attributes and their impact on reviews. However, successful models frequently employ black-box techniques that fail to explain how product attributes impact customer satisfaction. To address this, Explainable Artificial Intelligence (XAI) combined with advanced ML methods is suggested to create interpretable models. This paper evaluates the usefulness of combining LDA for topic modeling, Gradient Boost trees (GBT) for ML, and SHapley Additive exPlanations (SHAP) to develop an interpretable model for face mask attribute impact on online ratings. Analyzing 2,047 reviews of 35 mask products revealed seven key attributes. Labeling and nose-strap significantly influenced evaluations, while shopping experience, and packaging, filtering level, and fit were less impactful. This research supports the use of topic modeling combined with advanced ML techniques and XAI in analyzing online customer reviews to offer a time- and cost-effective method. It aids in understanding product attributes influencing satisfaction for product design and improvement, especially in the dynamic context of face mask preferences and purchasing decisions.