A Fast Autoencoder-based Recommender
Jiajia Jiang, Yunni Xia, Mingsheng Shang · 2019
High-dimensional and sparse (HiDS) matrices are the basic inputs of recommender systems. Recently, autoencoder-based approaches to analyzing HiDS matrices from recommender systems are becoming increasingly popular, owing to their good representative learning ability and scalability. However, traditional autoencoder-based approaches can usually bring great computational burden on HiDS data due to its frequent manipulations of large latent factor (LF) matrices data. To address this issue, we propose a Fast Autoencoder (FAE)-based recommender. It is capable of manipulating single LFs rather than LF matrices with improved computational efficiency. Moreover, we consider a Hogwild!-based parallelization mechanism for further accelerating its training efficiciency of the proposed recommender. Experimental results show that our proposed method considerably outperform traditional ones, e.g., the classical Autoencoder-based recommenders, in terms of recommendation accuracy and computational efficiency.