Parallelization of Latent Group Model for Group Recommendation Algorithm

Abstract

Recommendation system was proposed to solve the problem of information overload. Group recommendation is demanded as well as individual recommendation. Accuracy and efficiency come as main challenges in this field. Recently, group recommendation algorithm based on latent factor model has been proposed, which assumes that users are influenced implicitly by some latent factors. Existing method detects groups by considering latent factors and makes up users' profile in the form of latent factor. Then users' latent factor profiles were aggregated into a group profile and matrix multiplication was used for group recommendation. One of the core parts of this model is matrix factorization. Due to the high computational overhead of matrix factorization, it is relatively weak in big data processing. In this paper, we propose a Parallel Latent Group Model (PLGM) to improve the ability of processing large-scale data and to enhance the reliability and scalability. There are two models of matrix factorization in our consideration -- SGD and ALS. We implement parallel matrix factorization based on SGD on spark and compare it with ALS in MLlib. The strength and weakness of each model are analyzed based on the experimental result. Besides, different user profile aggregation strategies are studied in this paper and the best one is adopted to the model instead of the previous one. PLGM and LGM are compared in both accuracy and efficiency. Empirical studies on real datasets from MovieLens and Dianping.com demonstrate the effectiveness and efficiency of our improvement.