Research And Implementation Of Advertising Recommendation System Based On Network Representation Learning

With the development of the Internet and the popularity of various mobile devices,the amount of information in the network has exploded.Because of the stimulation of traffic realization,researchers gradually focus on the field of advertising recommendation.In the advertising recommendation,the number of users and the number of advertisements are increasing rapidly.How to provide personalized and precise recommendation for users from the massive advertising database based on the user portrait has become a common concern of research circles and industry.AD recommendation can be modeled as complex networks,where users and ads represent nodes in the network,and edges in the network represent relationships between node pairs.Intuitively,users'historical behavior information is crucial to user interest mining,and network representation learning is a learning method combining network node information and structure information.The current advertising recommendation system mainly has the following problems:(1)the advertising recommendation system mostly uses some artificial construction features as the original features to input the multi-layer neural network,which fails to well integrate the structural features and node features of the complex recommendation network scene,and the model training has artificial prior bias;(2)in complex network scenarios,nodes are sparse and obey power law distribution.How to deal with network sparsity and network information redundancy;(3)at present,the network-based representation learning algorithm is difficult to meet the needs of large-scale complex network scenarios,and there are few practices of large-scale network representation learning.In view of the above problems,this paper studies and practices the recommendation system based on network representation learning.The main work includes:(1)proposed a network construction method and an end-to-end training method in the context of advertisement recommendation.A hybrid higher-order learning framework(H-GAT)is proposed,and a hybrid higher-order sampling strategy is used to solve the problem of network construction for network representation learning.Based on the DSSM model,an improved model of graph aggregation two-tower model was proposed to train users and embed advertisements end-to-end.In the final online recall stage,Faiss library was used to accelerate the response time and accuracy of embedding retrieval.(2)a graph aggregation module VS-GCNs based on variational strategy and feature importance assignment is explored and implemented for network sparsity and cold start in advertising recommendation scenarios.Compared with the baseline model,the classification index of the benchmark data set of network representation learning is improved greatly.(3)study the implementation and optimization of the recommendation system based on network representation learning under distributed conditions.HiveQL and Hadoop streaming are used to accelerate network construction efficiency.GPU+PS cluster+DataFlow was used to accelerate the training of graph neural network.Finally,the recommendation system based on network representation learning is implemented in kuaishou advertising department.