| With the advent of the era of big data and the continuous development of artificial intelligence,the amount of data on the Internet is increasing exponentially.Hundreds of pieces of data are already beyond what users can handle,a phenomenon known as information overload.Therefore,the recommendation system has become an effective tool to solve the problem of information overload and has been widely used.The recommendation system can analyze the user’s preference information through the user’s historical interaction records,predict the user’s interest in the item,and recommend the most wanted content to the user.Especially when deep learning technology is introduced into the recommendation system,it can better understand the complex preference relationship between users and items,bringing opportunities for improving recommendation accuracy.Therefore,more and more researchers propose to integrate deep learning technology with traditional recommendation algorithms.Large-scale recommendation models based on neural networks usually have better performance due to their large capacity,and also have more parameters and longer inference time.Therefore,they mainly face two challenges in practical scenarios: On the one hand,these high-precision recommendation models are difficult to be directly applied to resource-constrained or real-time online recommendation platforms,and compressed recommendation models often suffer from obvious loss of accuracy.Therefore,how to achieve precision-controlled recommendation model compression is one of the challenges currently expected to be solved.On the other hand,in order to ensure the quality of recommendation,special recommendation models need to be built for different recommendation fields,and multiple similar recommendation models are maintained in multiple fields,resulting in increased maintenance costs and obvious waste of computing resources.Therefore,this paper focuses on the research on lightweight recommendation systems.For different recommendation scenarios,we will study recommendation methods that can ensure accuracy and achieve lightweight applications.The main research contents are as follows:（1）Aiming at the timeliness requirements of recommendation scenarios and maintaining the accuracy of recommendation results,this paper proposes a knowledge distillation recommendation method based on generative adversarial networks—GD-REC.On the one hand,the method aims to use the generative adversarial network to learn the potential preference features of users through adversarial training,and the generated recommendation list samples can alleviate the problem of poor model accuracy caused by data sparseness;on the other hand,the generated recommendation list samples and knowledge Distillation technology is integrated to train a recommendation model with excellent performance and few parameters,which improves the efficiency of the recommendation algorithm.The entire training process uses the generated recommendation list samples and the reconstructed knowledge distillation loss function to complete,that is,the final lightweight recommendation model can be trained without using the original data,enhancing the privacy protection of users.Finally,the experimental verification is carried out on the dataset of real scenes:compared with the comparison algorithm,the overall performance of this method can be improved by 7% on average,achieving a balance between utility and scalability.（2）In view of the problem of high maintenance cost and waste of resources caused by the existing methods in the dual-target cross-domain recommendation scenario for each field to build a recommendation model separately,this paper proposes a lightweight dual-target cross-domain recommendation method based on Star GAN: DT-Star GAN.This method can implement a generative recommendation model to provide high-performance recommendation services for users in two or more domains.Among them,the generator aims to generate recommendation list samples similar to user preferences,and the discriminator distinguishes the generated samples from the real interaction vectors in the domain.In addition,the Gumbel Softmax function is used to connect the generator G and the discriminator D to help the generator directly use the gradient descent method to accelerate the model convergence;a storage module is also added to the discriminator,which can more effectively help generate sample vectors and interaction vectors for dimensioning Alignment to improve the accuracy of the samples generated by the generator.Finally,to verify the effectiveness of the proposed model,this paper evaluates it on a real scene dataset.The experimental results show that the model and the comparison method in this paper improve the recommendation accuracy by 3%,and the convergence speed of the model and the design method are reasonable. |
