LeadRec: Towards Personalized Sequential Recommendation via Guided Diffusion

Sequential recommendation endeavors to forecast future interactions predicated on a user's historical behavioral sequence. The formidable expressiveness and high sample quality of diffusion models have unlocked numerous novel applications. Pioneering endeavors have showcased the efficacy of diffusion models in elucidating the underlying data patterns. However, existing methodologies tailoring diffusion models for recommendation tasks have not fully leveraged the sequence information and lack sufficient directionality when generating candidate items, resulting in suboptimal personalized recommendations. To address this disparity, we introduce an innovative Guided Diffusion Model framework dubbed LeadRec, designed to learn the generative process and make predictions in a guided manner. Specifically, LeadRec initially extracts a guiding signal from the historical interaction sequence using a transformer-based model. Subsequently, the sampled Gaussian noise undergoes iterative denoising by the guided denoiser. Ultimately, this denoising process yields a latent representation that encapsulates the user's interests and accurately reflects their true preferences. Extensive experiments and analyses conducted on three datasets corroborate the superiority of the proposed LeadRec framework.