Pretraining binarization encoders for recommendation acceleration
Feng Wei, Shuyu Chen, Yuan Yin, Hao Hu · Third International Conference on Artificial Intelligence and Electromechanical Automation (AIEA 2022) · 2022
Recommendations based on similarity search are widely used. However, when the number of users and items is large, it may become an efficiency bottleneck. In this paper, we propose a high-precision, modular, and efficient recommendation component: the binarization encoder. We use Siamese neural networks to jointly learn binary codes of users and items. The Siamese networks are composed of two autoencoders, whose encoder and decoder are sharing weights further. Their latent factors are discretized as efficient representations of users and items. In order to improve the learning ability of this model, a comprehensive objective function is proposed, which considers three objectives: improving the reconstruction quality of input vectors, reducing the information redundancy between dimensions, and finetuning the predictions with observed ratings. We use collaborative filtering based on matrix factorization on real-world datasets as basic systems and evaluate the proposed method. The experimental results prove the effectiveness of our method.