Research On Lightweight Sequence Recommendation Algorithm Based On Dynamic Combination Embeddin

At present,most novel and advanced sequential recommendation algorithms are all potential factor models,where each item is mapped to a single vector for representation to obtain the item embedding matrix.Then,the corresponding neural network is used to train the item embedding matrix and ultimately obtain the preferences of the target user through item embedding matrix.However,taking e-commerce platform as an example,there are inevitably tens of thousands of potential items in its system.If each item is mapped to a single vector in the embedding matrix one by one,when the sequential recommendation algorithms is deployed in a resource-constrained environment,the item embedding matrix will lead to memory bottleneck problem.In addition,many sequential recommendation algorithms are combined with some over parameterized neural networks,which leads to the problems of parameter redundancy and overcalculation in the training process,which affects the calculation speed and performance of the model.Therefore,lightweight sequential recommendation algorithms are urgently needed to be developed to solve these problems.Aiming at the memory bottleneck problem of previous item embedding matrix,this paper introduces the idea of compositional embedding,and designs a Dynamic Compositional Embedding algorithm DCE(Dynamic Compositional Embedding)to obtain the final item embedding matrix.Firstly,a smaller set of base embedding matrices are obtained for the interaction sequence through complementary partition.Then,a unique set of base embedding vectors are obtained from the base embedding matrices through Quotient-Remainder trick.Finally,the final item embedding vectors are generated by dynamically assigning weights to each base embedding vector and combining all base embedding vectors.The item embedding vectors generated by the Dynamic Compositional Embedding algorithm is more personalized and can more effectively represent the item,which is beneficial for the subsequent preference modeling.At the same time,the memory used to store the base embedding matrices is far less than the item embedding matrix in the potential factor model.In this paper,the effectiveness of DCE is verified by conducting sufficient experimental analysis on the publicly available datasets Movielens-20 M and Beauty.Aiming at the problem of parameter redundancy and overcalculation of previous sequential recommendation algorithms,this paper proposes a lightweight sequential recommendation algorithm DCELSRec based on Dynamic Compositional Embedding algorithm DCE in the preference modeling stage.DCELSRec introduces Dynamic Convolutional Neural Networks and Twin-head Self-attention Mechanism to extract users’ short-term and long-term preferences,respectively.Compared with the standard one-dimensional convolution,the order of training parameters required by the Dynamic Convolution Neural Network is far less than that of one-dimensional convolution.In the Multi-head Self-attention Mechanism,each attention head is used in parallel,which means that each attention head is calculated separately.Therefore,the more heads there are,the more unnecessary redundant parameters generated by the calculation,which will result in higher computing costs and memory costs.However,after sufficient experimental analysis,this paper adopts the Twin-head Self-Attention Mechanism with better performance.Based on the above research,DCELSRec has achieved lightweight features in the item embedding stage and preference modeling stage,making it a relatively complete lightweight sequential recommendation algorithm.This paper has set up sufficient experiments on three public datasets,Beauty,Yelp and Movie Lens-1M.The experimental results show that DCELSRec has improved in both two evaluation metrics compared with other advanced sequential recommendation algorithms,confirming the effectiveness of DCELSRec.