MAGE Bridges Multimodal Recommendation Gaps

MAGE tackles a real weakness in multimodal recommenders: similarity models often find substitutes while missing useful complements. Its geometry-plus-graph approach is promising, but the reported gains need careful scrutiny around co-purchase data and baseline parity.

• –Offline embedding warping gives complementary relationships a chance to influence downstream propagation.
• –Visual features remain stable anchors while augmented text features carry complementary signals.
• –The repository provides reproducible pipelines across four Amazon datasets, with pinned CUDA and PyTorch environments.
• –A key evaluation risk is whether co-purchase graphs expose information too closely related to held-out labels; stronger leakage controls would improve confidence.
• –The method is most relevant to e-commerce systems where accessories, bundles, and replacement items matter as much as visual similarity.