SADRec: Semantic-aligned adversarial diffusion model for multi-behavior recommendation

Multi-behavior recommendation leverages auxiliary behaviors as supplementary signals to identify the most commercially valuable latent target behaviors. While auxiliary behaviors provide valuable signals related to the target behavior, existing methods lack fine-grained alignment mechanisms between the target behavior and auxiliary behaviors. To address this limitation, we propose a Semantic-aligned Adversarial Diffusion Model for Multi-Behavior Recommendation (SADRec). Specifically, we first design an Adversarial Diffusion Model (ADM) with the target behavior as a discriminative anchor. ADM employs discriminative supervision to guide the reconstruction process, enabling auxiliary behaviors to adaptively identify key semantic dimensions and selectively preserve non-essential discrepancies when aligning with the target behavior, thereby achieving fine-grained and precise semantic calibration. Then, we introduce a Multi-level Negative Sampling mechanism (MNS) that extracts negative samples of multi-level hardness at specific time steps via the reverse process of the diffusion model and aligns them with the target behavior through multi-level graph contrastive learning, thereby jointly optimizing the reconstruction process and identifying auxiliary behavior signals that carry latent target behaviors. Finally, extensive experiments on three public multi-behavior datasets demonstrate that our SADRec outperforms state-of-the-art baselines, especially in alleviating data sparsity. Our datasets and source code are available https://github.com/sasdasdasdwew/SADRec.