A Memory-Efficient Approach to the Scalability of Recommender System With Hit Improvement

With the rapid increase of users and resources on the Internet, the scale of the recommender system becomes larger and larger. There are three major challenges facing in recommender system: sparsity, scalability, and cold start. In this paper, we mainly focus on the scalability issue and propose a recommender system based on the memory-efficient recurrent neural network. First, we allocate an item table for items and use a pair of embedding vectors to represent each item. Thus, we can use a few vectors to represent numerous items and decrease the memory used for the storage of embedding vectors. Second, we present a similarity-based initialization method for the item table to get a better representation of items. Third, we further design the loss function and the adjustment method to adjust the placement of items in the item table to speed up the training procedure of the model and get a better performance. The experimental results demonstrate the effectiveness of our approach. It can clearly improve the performance of recommendation, such as hit rate and normalized discounted cumulative gain, when compared to the state-of-the-art recommender algorithm. In addition, our approach can also handle the cold start problem and supply new users with the same quality of service as the old users.