| 5G messages integrate a variety of rich media formats through the native SMS entry,realizing SMS as application and SMS as service.With an open attitude,5G news strives to build short messages into a unified business entrance,and pushes messages without requiring users to pay attention to the service number.However,as the number of 5G message service and the amount of 5G message data increase,blindly pushing 5G message to users will inevitably lead to overrecommendation and low user satisfaction.Therefore,how to push the message that users need has become an urgent problem to be solved.In order to improve user experience and user satisfaction,5G message recommendation system needs to generate message recommendation schemes for different users based on user portraits and message entities.By using the user’s behavior sequence and the content of the message entity as the two main data characteristics and combining the contentbased and collaborative filtering recommendation methods,satisfactory 5G message recommendation results can be generated for users.Therefore,it is of great significance to study the 5G message recommendation method based on user behavior portrait.Firstly,5G messages contain a variety of data,such as images and texts,and the simultaneous use of image and text features can better characterize 5G messages.Therefore,this paper proposes a 5G message session recommendation model based on feature fusion.This model can better characterize the features of 5G messages by extracting and integrating the image and text features of 5G messages.In addition,the model improved the ranking loss function to optimize the classification results.Finally,the model uses a mini-batch training mechanism to add additional negative samples,which effectively improves the recall rate.Then,considering the problems of cold start and matrix sparsity in traditional collaborative filtering methods,this paper designs a meta-level collaborative filtering recommendation method based on user clustering after comparing optimization methods of various recommendation systems.Based on the static information and dynamic behavior information of users,the user portrait is constructed.The local linear embedding algorithm and the improved carnivorous plant algorithm are used to cluster users.Then,the initial recommendation results are generated according to similar users clustered by users.Finally,the second filtering of the recommendation results is carried out through the memory time window mechanism,which effectively alleviates the problems of cold start and matrix sparsity,and improves the efficiency of calculation and the accuracy of recommendation.Finally,simulation experiments are carried out on the two proposed methods,and the two methods are implanted into the actual 5G message recommendation system,and the final recommendation results are generated through the mixed recommendation of the two types of recommendation items set.The feasibility and accuracy of the method are proved by the experimental results and the actual operation of the system. |
