A Meta-Learning-Based Solution to Address the Sparsity Problem of Recommender Systems
Zhi Chai, Cheng Zhong Yang · 2019 International Joint Conference on Information, Media and Engineering (IJCIME) · 2019
Sparsity is an important challenge for recommender systems. Most of the existing solutions to sparsity problem are achieved by adding auxiliary data. However, in practical recommender systems, auxiliary data is often difficult to obtain and limited in number. Considering the characteristics of data sparsity, this paper transforms it into a meta-learning problem. We propose a new meta-learning solution based on multi-layer perception (MLP) network to solve the sparsity problem of recommender systems from the perspective of optimizing parameters with a network using meta-learning. We split the behavior data into small sets by users and consider the training process of a small set as a task, and train an MLP network through these tasks to learn a parameter-optimizing strategy of the recommendation model. The network can quickly fit the extremely sparse data and optimize recommendation model parameters. We then predict the recommendation results on the test set with the optimized model parameters. Experiments on movielens and NetFlixPrize datasets show that the Root Mean Square Error (RMSE) value of the reference recommendation model with meta-MLP increases by 1%-10% when the data sparsity is less than 0.001%.