Diffscs: improving strict cold start recommendation with diffusion model

Strict cold start recommendation poses a significant challenge to recommender systems. It results in traditional collaborative filtering methods being ineffective due to limited or no historical interactions available to new users/items. However, there is little academic research on this topic. Therefore, we propose a novel strict cold start recommendation approach based on the diffusion model, called DiffSCS. Firstly, we introduce a novel embedding generation framework comprising a content-embedding generator and a diffusion-embedding generator. The content-embedding generator captures high-order interactions among content features using Bi-Interaction pooling operations. The diffusion-embedding generator produces preference embeddings for new users and items based on their content embeddings employing a diffusion model. Then, we introduce a neighbor selection GCN that utilizes attribute similarity and rating diversity as the benchmarks to guide contextual neighbor selection on attribute graphs for each user/item to ensure that the chosen neighbors are more representative. Next, we integrate the preference and content information of the new user/item and that of its selected neighbors to enhance the final embeddings of the new user/item. Finally, a large number of comprehensive experiments are performed on three common datasets with seven contrasting algorithms. Experimental results show that compared with the best comparison algorithm, DiffSCS can still improve by 0.60% \(\sim \) 8.31% in strict cold start scenarios which demonstrates the superiority of DiffSCS.