Style-aware digital media art recommendation based on multimodal big data and graph learning

Recommending digital media artworks to users is more difficult than recommending general products for two reasons: user preference depends on artistic style as well as content, and interaction records are typically sparse. Existing multimodal recommendation methods treat visual features and text embeddings as generic item descriptors, without explicitly separating style as an aesthetic dimension or matching it to individual user history. We propose SAGMR (Style-Aware Graph Learning for Multimodal Recommendation), which maps visual, textual, and style features into a shared latent space, weights modality contributions through an item-specific attention mechanism conditioned on each artwork’s style representation, explicitly matches candidate styles against each user’s historical style distribution, and propagates collaborative signals through a lightweight user–artwork bipartite graph. The final ranking score combines graph-learned collaborative preference with a direct style matching term. Experiments on the Amazon Baby benchmark, used as a cross-domain proxy with structurally compatible sparsity and multimodal features, show that SAGMR outperforms seven baselines on Recall@K and NDCG@K at two cutoff values, with the largest gains for users with fewer than five training interactions. Ablation results confirm that graph propagation and textual features are the two largest individual contributors.