基于动态组合嵌入的轻量级序列推荐算法

Dynamic Compositional Embedding Based Lightweight Sequential Recommendation

现有的大多数序列推荐算法,将每个项目映射到一个向量进行表示,当项目数量过于庞大时,其项目嵌入表存在内存效率问题;另外很多序列推荐算法与一些过度参数化的网络相结合,导致训练过程中存在参数冗余的问题,影响模型的计算速度和性能。针对以上问题,该文设计了一种轻量级的序列推荐算法,以取得相比以往方法更高的内存效率。首先使用动态组合嵌入方法,通过互补分区生成一组更小的基嵌入表,并使用商余技巧和权重分配动态地生成最终的项目嵌入。其次,为了避免参数冗余,引入动态卷积网络和双头自注意力来提取用户的短期和长期偏好。结合以上两个部分,得到一种轻量级的序列推荐算法DCE-DCN,并且在三个公开数据集Beauty、Yelp和MovieLens-1M上设置充分的实验验证了算法的有效性。

Most of the existing sequential recommendation algorithms map each item to a vector for representation,which suffers from a memory efficiency problem in its item embedding table amd a parameter redundancy issue in the training process.This paper designs a lightweight sequential recommendation algorithm to achieve higher memory efficiency.Firstly,the Dynamic Compositional Embedding method is used to generate a set of smaller base embedding tables through complementary partitions,and the final item embedding is dynamically generated by Quotient-Remainder trick and weight allocation.Secondly,Dynamic Convolution Network and Twin-head Self-Attention are introduced to extract users'short-term and long-term preferences to avoid parameter redundancy.Combining the above two parts,a lightweight sequential recommendation algorithm DCE-DCN is obtained,and experimental results on three public datasets Beauty,Yelp and MovieLens-1M are reported.