Modern e‑commerce platforms naturally provide three complementary modalities: textual item descriptions, item images, and user reviews. Existing multimodal recommendation methods often omit reviews or fail to integrate all three modalities within a unified framework. To address this gap we introduce MPIUR, a multimodal graph‑based recommendation model. The model first constructs a heterogeneous graph from user reviews, applying graph neural networks to capture semantic relations among reviews. Then a shared projection layer aligns image features with text features, bringing visual and linguistic representations into a common latent space. To handle varying item interaction density, MPIUR incorporates a collaborative‑semantic alignment mechanism that adaptively balances collaborative filtering signals and semantic similarity based on sparsity. Experiments on Amazon Reviews 2023 Musical Instruments and Baby categories show that MPIUR achieves the best global nDCG@10 and Recall@10 among all baselines, and also excels on sparse items, demonstrating that jointly exploiting reviews and multimodal item content is especially beneficial when interaction histories are limited.
Review: MPIUR’s heterogeneous‑graph encoding of reviews combined with cross‑modal alignment markedly improves recommendation performance in sparse scenarios, offering a scalable unified framework for multimodal recommendation.