Research On Privacy-preserving E-commerce Intelligent Recommendation And Lightweight Algorithms

With the continuous development of e-commerce platforms,how to help users quickly discover items of interest from a massive inventory has become an important factor for platforms to achieve more profits and long-term development.In this process,designing efficient sequence recommendation algorithms by collecting and analyzing user behavior data is a key means for platforms to improve user satisfaction and stickiness.However,current sequence recommendation algorithms often only consider the order information of user behaviors when modeling user behavior sequences,ignoring the fine-grained time characteristics of user behaviors,which makes it difficult to accurately capture user needs and preferences at different times and affects the effectiveness and quality of recommendations.On the other hand,with the awakening of users’ privacy protection awareness and the legislation by the government to restrict service providers from collecting user data,recommendation systems need to pay more attention to users’ privacy protection needs.The emergence of federated learning provides a solution for privacy protection in recommendation systems.However,in the federated learning scenario,the storage and computing capabilities of user-side devices participating in federated learning vary,and in resource-limited devices,they may not be able to bear the computation and storage load required by recommendation models,which affects the deployment and operation of models locally.In response to the above issues,this paper conducts research in three aspects of the recommendation system: privacy protection requirements,user behavior sequence modeling,and model lightweight design.The main contributions and innovations of this paper are as follows:(1)In the aspect of privacy protection in recommendation systems,this paper proposes a privacy-preserving e-commerce intelligent recommendation service architecture.The architecture fully respects users’ privacy and provides users with the choice of data collection and use.It offers two modes of recommendation services: traditional centralized recommendation service and federated learning-based recommendation service.For users who agree to share their data,their data will be used for centralized recommendation model training and usage through a data access service.For users who do not agree to share their data,they can train and use a lightweight recommendation model locally through the federated learning module,thus achieving user-controlled privacy protection.(2)In the aspect of user behavior sequence modeling,a convolutional sequence recommendation algorithm that integrates self-attention and fine-grained time features is proposed.This algorithm fully utilizes the fine-grained time features of user behavior,captures the long-term interests and short-term preferences of users in the sequence through self-attention mechanism and convolutional operation respectively,and improves the efficiency of utilizing time information in sequence recommendation models.Experiments on multiple datasets have shown that compared to several baseline models,this algorithm can effectively improve the accuracy of recommendations.(3)In the aspect of model lightweighting,a lightweight hybrid sequence recommendation model is proposed.This model fully utilizes user-item rating information,adopts pre-training to reduce the huge parameter size caused by end-to-end user embedding training for the model,and uses a self-attention mechanism with improved computational efficiency.In addition,the model reduces storage overhead through self-attention distillation.Experimental results show that the improved lightweight model has fewer parameters and lower computational complexity than the original model,making it more suitable for deployment on user devices with limited resources in a federated learning scenario.