NMF-based DCG Optimization for Collaborative Ranking on Recommendation Systems

Abstract

A recommendation system predicts a top-N list of items that a target user might like by considering the user's previous rating history. In this paper, we solve the task of recommendation by developing a method that implements an NMF-based DCG optimization for collaborative ranking. Three main processes are applied to calculate the rating prediction for making the list of top-N item recommendations: constructing the user profile, initialising the latent-factor models using NMF (Non-Negative Matrix Factorization), and further optimising the models based on the DCG (Discounted Cumulative Gain). Extensive evaluations show that our proposed method beats all baseline methods on both the Precision and NDCG metrics. This fact confirms that NMF-based DCG optimization is an effective approach to enhance the recommendation performance and to deal with the sparsity problem.