Research On Deep Recommendation Algorithm Based On Graph Embedding For Information Enhancement

In the information age,the problem of information overload has exacerbated the widespread application of recommender systems in various online services.In order to improve the recommendation accuracy,deep recommendation models have been widely deployed in practical applications to generate more useful recommendation results for users due to their better expressiveness.At the same time,rich side information is also widely used in the deep recommendation model to generate information-enhanced recommendations to further improve the recommendation effect.However,the current deep recommendation algorithms cannot integrate side information efficiently and flexibly due to the limitations of scalability and expressiveness in the feature embedding layer,which are seriously affected by the feature input dimension.Based on the above problems,this thesis starts research on how to efficiently utilize side information by graph embedding for recommendation to provide more satisfying recommendations to users.The details of the research are as follows.(1)A deep recommendation architecture based on graph embedding for information enhancement is proposed,which separates representation learning and task learning to flexibly incorporate various types of side information.Firstly,the side information is represented in the form of a graph,and then the graph embedding algorithm is used for representation learning on the graph data to obtain a feature vector with rich information,that is,the embedding of the node,and then the feature vector is used as the input of the deep algorithm as the task learning.In addition,in view of the inconsistency of the feature space of the feature vector generated by graph embedding,the deep recommendation architecture is adjusted,and multi-tower layer are added to alleviate the problem of the inconsistency of the feature space and mine the information in the feature vector fully;Aiming at the problem that key business features are easily annihilated by deep feature interaction,an FM structure is added to "remember" key business features.Finally,experiments based on real-world datasets verify the effectiveness of the proposed model,the influence of different graph embedding algorithms and multi-tower layer,and the computational cost advantage of the proposed model compared to traditional recommendation algorithms.(2)Aiming at the cold start problem of embedding that the model proposed in this thesis may encounter in practical applications,the solution to the cold start problem is studied.This thesis proposes solutions from three perspectives: cold start supplementary mechanism,graph embedding algorithm and deep architecture adjustment.Based on two real datasets,different degrees of cold-start datasets are constructed,and the proposed cold-start solutions from different angles are tested on these datasets.Firstly,combined with the experimental results,the sensitivity of different cold-start solutions under different cold-start degrees and the reasons are analyzed.In addition,the comprehensive performance of different cold-start mechanisms under different cold-start degrees is compared and analyzed.Finally,combined with the experimental results,the applicability of different cold-start solutions in different application scenarios is discussed to guide the practical application.This study provides a novel solution to efficiently utilize side information to generate information-enhanced recommendations,enriching research on feature-pretrained deep recommendation methods beyond end-to-end recommendation architectures.In terms of application value,this research provides a model architecture for online recommendation services in the real world that takes into account both recommendation quality and service efficiency,which is beneficial to the platform’s user retention and profit acquisition.