Improved spectral clustering collaborative filtering with Node2vec technology

Recommender systems are becoming part of our lives and make many online systems getting smarter. However, with the fast increasing amount of users, items, and the interactions between the two, many traditional recommendation methods encounter the challenges of sparsity and efficiency. Clustering-based recommendation is a common strategy to overcome these problems, but many of them rely heavily on parameter settings. In this work, we put forward an improved spectral clustering-based collaborative filtering framework based on node2vec algorithm, named as N2VSCDNN, to tackle these challenges and lead to a significantly increment of accuracy. We represent interaction data as a bipartite network and employ node2vec technology to automatically capture complex latent relationship among users or items from the corresponding projection networks. In order to overcome the information loss problem and reduce the dependency on parameter settings, We proposed an improved parameter-free spectral clustering based on dynamic nearest-neighbors to cluster users and items separately. After partitions of the nodes, we generate recommendation over the most frequent pairwise user-item clusters and merge the results to conduct a top-N list for every individual user. Finally, We demonstrate the effectiveness of our proposed framework on a real-world benchmark dataset Yelp.