Research And Application Of Recommender Systems Based On Graph Embedding Techniques

In today's “fourth technological revolution” of mankind,physics is driving the development of artificial intelligence in a unique way.As one of the prevalent applications of artificial intelligence,recommender systems aim to automatically predict user's interests or most needed items from millions of candidates,alleviating the information overload problem.At the same time,as an interdisciplinary research field of physics and sociology,network science is being increasingly penetrated into the research of recommender systems,used to promote recommender systems' development and breakthroughs.One typical combination between them is graph embedding-based recommender systems.At present,the research on graph embedding techniques lies in adopting machine learning methods.However,since the rationale behind these methods is based on the“black box” mechanism of fitting optimization theory,most graph embedding-based recommender systems are inevitably facing fundamental limits like weak interpretability or hyperparameter uncertainty.Possible ways out of this dilemma were to establish a theoretical connection between model's hyperparameters and its performance from a mathematical perspective or to take machine learning methods as a bedrock to incorporate more interpretable elements.Admittedly,these explorations to some extent contributed to alleviating the above difficulties.However,they still cannot fundamentally solve the problems of model's weak interpretability or decrease model's hyperparameter uncertainty due to their adoption of machine learning frameworks.Consequently,in practical recommendation systems,enterprises would have to spend enormous computing resources to work out a set of hyperparameters that can bring good performance for recommendation models,and business personnel would still find difficulties in making reliable explanations for recommendation results generated by these models.To this end,from another research perspective,by combining the field of network science and the essential thoughts(not the specific frameworks or models)of machine learning methods this thesis proposes a hyperparameter-free and interpretable whole graph embedding technology as well as a hyperparameter-free and interpretable recommendation model.In detail,the main contributions of this thesis include:1.This thesis gives a comprehensive review of recommender systems,comparing the pros and cons of different recommendation methods in theory,and proposing a general design pipeline of graph embedding-based recommendation models.2.This thesis designs and implements a comparison experiment of classic recommendation methods,comparing the pros and cons of conventional recommendation methods and graph embedding-based recommendation methods in practice,and proposing strategies for making a trade-off between them.3.This thesis proposes DHC-E algorithm,a hyperparameter-free and interpretable whole graph embedding technique.Compared with four baselines,DHC-E can achieve the overall best accuracy and stability,as well as good computing efficiency in binary graph classification tasks and multiple graph classification tasks.4.This thesis proposes AIProbS model,a hyperparameter-free and interpretable recommendation model.Compared with nine baselines,AIProbS can achieve the best accuracy and good stability on some recommendation tasks.