Research And Application Of Multi-Scenario Recommendation Based On Graph Neural Network

With the continuous development of information technology,the Internet has become an integral part of people’s daily life.At the same time,the problem of information overload is becoming more and more important.It is difficult for users to quickly locate their needs in a large amount of data.The emergence of recommendation system is to solve the problem of"information overload".For example,product recommendation can provide accurate product recommendation to users by deeply mining user preferences.At the same time,along with many complex scenarios in practical problems,recommendation tasks have also become diverse,and the user-oriented recommendation task model and the item-oriented recommendation model have also gradually.In recent years,the rapid development of graph neural networks has provided a strong foundation and opportunity for solving the above problems in recommendation systems.Specifically,GNN iteratively aggregates neighborhood information by spreading embedded information.Through the stack propagation layer,each node can access the information of its higher-order neighbors,instead of only accessing the information of its first-order neighbors as the traditional method.With its unique advantages in processing structural data and exploring structural information,GNN-based methods have become the most advanced new methods in recommendation systems.In academic research,a lot of work has shown that the model based on GNN is superior to the previous methods,and has made a lot of latest achievements on the public benchmark data set.In industry,GNN has also been deployed in large-scale recommendation systems to produce high-quality recommendation results.In order to fully demonstrate the effectiveness of GNN in multiple recommendation scenarios,this paper analyzes the usability and defects of GNN in specific scenarios from the three modes of "multi-behavior recommendation","bundling recommendation" and "pushing things by things",and successively completes the design and demonstration of graph models in different recommendation scenarios.In order to explore the "multi-behavior" recommendation model in ecommerce scenarios,this topic first analyzes the advantages of multibehavior recommendation over single-behavior recommendation from the data level,and analyzes the inherent prior knowledge of the recommendation scenario after the introduction of multi-behavior.Secondly,in order to achieve "seeking common ground while reserving differences" among multiple behaviors,we designed a multi-behavior recommendation model S-MBRec based on graph self-monitoring,designed an adaptive semi-monitoring task and a star self-monitoring task,and the two tasks were jointly optimized.Finally,a large number of experiments on public data sets show that our proposed model can effectively improve the recommendation performance under multiple behaviors.In order to explore the problem of bundled recommendation under commodity recommendation,this topic is based on the scenario of China Mobile’s equity package recommendation,and it is analyzed that the recommended package is in the combined state(i.e.the combined package of multiple APP).The ternary relationship between "user-package-APP",that is,the user and the benefit package have interactive information;Secondly,the package and APP also have interactive information,indicating which kinds of APP are included in each package;Finally,there is also interaction between users and APPs to indicate which APPs users have used.Under this research,we first analyzed the business value and logic behind the equity package recommendation from shallow to deep,and then carried out data cleaning and feature engineering for real scene data.Finally,the corresponding graph model design scheme is given to generate and mine the comprehensive preferences of users in the package and APP dimensions,generate the recommendation results,and finally deploy the model to the internal platform of the mobile company.In order to explore the recommendation model of "pushing things by things",this topic starts from the practical application of similar store detection(input a store data,return the most similar stores,and recommend it to managers)selected by Meituan,and deeply excavates the modeling characteristics and unique challenges of pushing things by things.The purpose of this research is to explore the repetition&similar entity recommendation(retrieval)under the graph neural network,and think about "how to explicitly model the association between entities?How to describe the association between multiple modes?How to solve the challenge of extreme lack of labels in real scenes?".In this research,we use multimodal features to construct a multi-relationship diagram,and then design an effective comparative learning and self-training learning framework to deal with the problem of extreme lack of labels.Finally,a set of graph contrast self-training model(CT-GNN)is generated.A large number of experiments have proved the effectiveness of our proposed model.