Exploring lottery ticket hypothesis in media recommender systems

• DOI: 10.1002/int.22827
• Publication Date: 2022-01-17T09:37:50Z
• Authors: Yanfang Wang; Yongduo Sui; Xiang Wang; Zhenguang Liu; Xiangnan He
• Keywords: Lottery; Ticket

Media recommender systems aim to capture users' preferences and provide precise personalized recommendation of media content. There are two critical components in the common paradigm modern models: (1) representation learning, which generates an embedding for each user item; (2) interaction modeling, fits toward items based on their representations. In spite great success, when a amount users exist, it usually needs create, store, optimize huge table, where scale model parameters easily reach millions or even larger. Hence, naturally raises questions about heavy Do we really need such large-scale parameters? We get inspirations from recently proposed lottery ticket hypothesis (LTH), argues that dense over-parameterized contains much smaller sparser sub-model can comparable performance full model. this paper, extend LTH systems, aiming find winning tickets deep models. To best our knowledge, is first work study systems. With Matrix Factorization Light Graph Convolution Networks as backbone models, found there widely exist On three convergence data sets—Yelp2018, TikTok Kwai, achieve with only 29 % ~ 48 , 7 10 %, 3 17 parameters, respectively.