Research And Application Of Feature Intersection In Advertising Recommendation Algorithm

Advertising recommendation algorithm is an important research field in the development of the Internet.Its continuous progress enables many more valuable information to be pushed to users with different interests and preferences,which improves the efficiency and quality of our daily information acquisition.During the development of advertising recommendation algorithms,many new technologies have emerged,among which feature intersection is an important technology.Feature intersection is a high-level representation of feature synthesis.The purpose of feature intersection is to obtain nonlinear features,increase the information input of the model,and improve the representation ability of the model.Existing studies have improved the performance of the model by designing a better feature intersection algorithm,but there is still a problem that the standard deviation of the results of multiple rounds of inference is large.In order to solve this problem,this paper proposes a Multi-learner Network named MLN,which is based on the idea of ensemble learning.The MLN model mainly includes a reweighting layer,a multi-learner layer and an ensemble layer.Through feature reweighting,network integration and self-knowledge distillation,the model performance exceeds some cutting-edge research while reducing model variance.In the data preprocessing part,in order to alleviate the problem of long-tailed distribution of continuous values,this paper proposes a nonlinear transformation preprocessing scheme.In order to solve the problem of sparse ID feature distribution,this paper proposes a frequency division dictionary encoding scheme.The experimental results show that the proposed preprocessing scheme has better performance in benchmark experiments.Finally,this paper analyzes the system requirements in order to designs and implements an advertising recommendation system.The system is based on big data and deep learning related frameworks,including log storage service,feature service,model platform,recommendation service and unified service platform.It can complete endto-end advertising recommendation.The proposed preprocessing methods and MLN model in the paper are applied to this system after functional testing and performance testing.