Byungmoon Heo

Sequential recommendation (SR) predicts a user's next action by modeling their historical interaction sequence, and it underpins many real-world platforms where accuracy directly impacts engagement and conversion. Recently, diffusion-based SR has gained traction by iteratively denoising corrupted representations, improving robustness and handling uncertainty when user intent is ambiguous. However, most diffusion-based SR methods still rely mainly on item IDs and learned latent embeddings, complemented by conditioning or regularization techniques. This ID-centric design limits semantic expressiveness: semantically similar items can remain distant under sparse feedback, while unrelated items may be confounded by co-occurrence and popularity biases. Prior work has attempted to inject semantics through contrastive learning or auxiliary objectives, but semantic signals are often incorporated indirectly and can weaken along the denoising trajectory, especially at inference when generation starts from noise. To address this limitation, we make semantics an explicit denoising target by introducing a semantic token whose target is the next-item semantic embedding in the same latent space as item embeddings, and we jointly denoise the item sequence and the semantic token. Building on this idea, we propose SEDRec (Semantic Entanglement generative Diffusion for sequential Recommendation), which entangles semantic and ID-based signals through history semantic conditioning, synchronized noise levels, semantic-to-target mixing, and intermediate alignment to a fixed semantic anchor. Across multiple benchmarks and evaluation metrics, extensive experiments demonstrate that SEDRec consistently outperforms strong baselines, and ablation studies confirm the contribution of each semantic entanglement component.