What makes a review helpful? A multimodal prediction model in e-commerce

Abstract

The rapid expansion of e-commerce has made online reviews critical to consumer decision-making. However, the overwhelming volume and inconsistent quality of these reviews significantly complicate the identification of valuable information. Consequently, machine learning and deep learning-based review helpfulness prediction (RHP) has emerged as a robust approach for addressing these challenges. While unimodal RHP methods primarily analyze textual content, multimodal approaches enhance predictive accuracy by integrating textual and visual content. Despite these advantages, existing models often overlook shallow features, such as review readability and image quality, which are critical to consumer evaluation. To provide a more comprehensive framework for RHP, an elaboration likelihood model–grounded multimodal cue-based helpfulness-prediction model (MCHPM) is proposed. MCHPM extracts central cues (semantic content from text and images) using bidirectional encoder representations from transformers (BERT) and VGG-16, whereas Python-based feature extraction captures peripheral cues (e.g., text readability and image quality). Finally, a co-attention mechanism identifies cue interdependencies, and a Gated Multimodal Unit effectively weighs each modality during prediction. Empirical evaluation on a real-world Amazon review dataset demonstrates that the proposed MCHPM outperforms benchmark models, thus validating the criticality of incorporating shallow features into multimodal RHP studies. Overall, these findings offer valuable insights into enhancing RHP and ensuring more informed consumer decision-making in e-commerce.