GCD-SPKAU:A Model-based Meta-learning with Sharing Prior Knowledge Among User-clusters classified through Graph Community Detection for Cold-Start Recommendation

Abstract The cold-start problem has consistently been a significant challenge in recommender systems, particularly in the initial stage for new users or new items. This is due to the lack of sufficient historical interaction data, which presents a significant obstacle to the development of effective personalized recommendations. Nowadays, a considerable number of researchers and scholars have investigated the issue of cold-start problems, resulting in the emergence of a multitude of sophisticated methodologies aimed at addressing this challenge. Among these, meta-learning has emerged as a particularly promising avenue of enquiry. Most existing meta-learning algorithms assume that a priori knowledge can be shared globally among all users. However, a problem arises that information sharing among users with widely different interests is usually ineffective, which leads to suboptimal solutions. Therefore, it is important to improve the meta-learning model. As a result, this paper proposes a new, improved meta-learning model idea, which first invokes the clustering algorithm of Graph Community Detection to classify different categories of user clusters, and then obtains the unique prior knowledge belonging to themselves in each user cluster, and finally enables the prior knowledge to be shared among the clusters of those categories that have the most effective effects. We compare the performance advantages and disadvantages of the new model and the traditional model in cold-start recommendation, verify the effectiveness of our new idea in improving the performance of cold-start recommendation, and discuss its potential applications and future research directions.