基于HashNet的深度学习推荐系统嵌入表压缩方法  

Deep Learning Recommendation System Embedding Table Compression Method Based on HashNet

作　　者：潘志安[1] PAN Zhi-an(School of Information Engineering,Hubei Polytechnic Institute,Xiaogan,Hubei 432000,China)

机构地区：[1]湖北职业技术学院信息工程学院,湖北孝感432000

出　　处：《湖北职业技术学院学报》2022年第4期95-100,共6页Journal of Hubei Polytechnic Institute

摘　　要：如何将用户感兴趣的信息进行精准推送是一项重要的研究课题。目前的通用方法是将类别信息训练为稠密的矩阵形式,这种矩阵也被称为嵌入表(embedding table),这些嵌入表通常会占据极高的内存空间,导致模型的可拓展性大幅降低。为解决这一问题,提出了HashNet Embedding(HN embedding)方法,通过矩阵单元共享的方式,对嵌入表进行压缩。在Criteo数据集上的实验表明,HN embedding方法可以在几乎没有性能损耗的情况下,将嵌入表内存压缩为1/32。How to accurately push the information that users are interested in is an important research topic.The current method to deal with sparse features in recommendation is to train category information into a dense matrix form,which is also known as"embedding table",but these embedding tables often take up a very high amount of memory space,resulting in a significant decrease in the scalability of the model.To solve this problem,a HashNet Embedding(HN Embedding)method is proposed in this paper.HN Embedding can compress the embedding table by sharing the cells in the embedding table matrix.Experiments on the Criteo dataset have shown that the above-mentioned HN embedding method can compress the embedded table memory to 1/32 without any performance loss.

关 键 词：推荐系统 点击率预测 深度学习 嵌入表 内存压缩 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]