Abstract
In recommender systems, new items cause the Cold-Start problem due to the lack of user interaction data, making collaborative filtering difficult to apply. While generative models leveraging side information, particularly Diffusion Models, have gained attention as a solution, their iterative denoising process remains a structural bottleneck that limits real-time service deployment. To address this, we propose a side information-based Conditional Flow Matching recommendation model that resolves the computational bottleneck of Diffusion Models and improves inference efficiency. The proposed model learns trajectories that directly predict target user interaction vectors conditioned on item content, enabling meaningful recommendations even for new items without interaction history. By adjusting inference steps, the model achieves at least 10× faster inference compared to the Diffusion Model. This demonstrates a practical trade-off between recommendation quality and computational efficiency, confirming that the proposed model is a more suitable and scalable methodology for large-scale, real-time Cold-Start recommendation environments.