Multimodal Deep Learning for Online Review Helpfulness Prediction

ABSTRACT Review helpfulness prediction (RHP) is critical for alleviating information overload and supporting reliable decision making on e‐commerce platforms, yet it is challenged by unstructured and multimodal review content as well as distorted helpfulness signals. Prior studies have examined textual, structural, and reviewer‐related determinants and built models based on manually engineered features or deep learning, but most approaches remain text‐centric, make limited use of review images and structured metadata, and offer little interpretability regarding how different information sources contribute to predictions. To address these issues, we propose Attentive Gated Multimodal Fusion (AGMF), a deep learning framework that jointly models review text, review images, and structured metadata. AGMF employs a pre‐trained language model for textual representations, a convolutional neural network for visual features, and metadata features capturing reviewer characteristics and behavioural signals, and fuses them through cross‐modal attention and a gated multimodal unit that adaptively weights each modality at the instance level, providing modality‐level interpretability. Experiments on a large‐scale dataset of Chinese e‐commerce reviews show that AGMF consistently outperforms traditional machine‐learning methods, strong single‐modality deep learning baselines, and competitive multimodal fusion models in terms of accuracy, F1‐score, and AUC, and ablation studies confirm the effectiveness of each modality and the proposed fusion mechanism. Overall, this study contributes an interpretable multimodal architecture for RHP that effectively integrates text, images, and metadata and offers practical guidance for designing intelligent review filtering systems on online platforms.