Graph Neural Networks For Personalized Recommendation

The rapid development of Information Technology and E-commerce prompts the popularization of the Internet,which has now become an indispensable part of people’s daily life.While the massive information on the Internet enriches people’s lives and meets people’s growing material and spiritual needs,it also poses challenges of(1)presenting information that meets personalized tastes for users,and(2)delivering information accurately to audiences for platforms.As an effective tool to alleviate the problem of "information overload" in the era of big data,personalized recommendation has become one of the key technologies to support the smart Internet and the high-end smart economy.The core of personalized recommendation is modeling user intent.Traditional heuristics-and matrix factorization-based methods can only model shallow and implicit user intent,while fall short in capturing multiple high-order interaction patterns in an explicit way.Given that(1)most data in recommendation can be organized using graph(e.g.,user-item interaction data,users’ social networks,items’ knowledge graphs,etc.),and(2)graph neural network has shown its great power in modeling high-order complex relationships,graph-based recommendation has drawn extensive attention from researchers.However,there are many shortcomings in the existing work.For example,at the recommendation modeling level,the ability of existing models for modeling graph data is somehow weak.On the one hand,it’s difficult to fully exploit the unlabeled data space,while on the other hand,it cannot simultaneously capture multiple interaction patterns in the heterogeneous data.At the model optimization level,the existing work often adopts classical loss function for model optimization,lacking the adaptation to the characteristics of the graph-based methods and in-depth theoretical analyses,to name a few.In response to these challenges,this paper carries out studies on both recommendation modeling and model optimization.Specifically,for recommendation modeling,this paper innovatively proposes to enhance the graph-based recommendation from two aspects:endogenous graph data modeling and exogenous graph data modeling;while for model optimization,this paper offers a comprehensive understanding of the optimization mechanism for the graph-based recommendations.The main contributions and innovations are as follows:First,graph-based recommendation augmented by endogenous data:Aiming at the problems of sparse supervision signal,skewed data distribution,and noisy interaction data that existed in the prior graph-based recommendation algorithms under supervised learning paradigm,this paper focus on investigating self-supervised learning in recommendation,by exploiting the unlabeled data space.The idea is to set an auxiliary self-supervised learning task to complement the classic supervised recommendation task.Specifically,we develop five augmentation operations to construct the unlabeled data space by changing either ID embeddings or graph structures,where each operator works with different rationality.The objective of self-supervised learning is to maximize the agreement between different augmented views of the same node.Furthermore,we prove in theory that the proposed method inherently encourages learning from hard negatives.Extensive experiments on benchmark datasets justify the advantages of the proposed method regarding both normal and long-tail recommendation tasks,training efficiency,and robustness against noisy interactions.Second,graph-based recommendation augmented by exogenous data:In view of the multi-source heterogeneous data involved in recommendation,and the defects of inability to capture simultaneously multiple interaction patterns and low inference efficiency in the existing context-aware recommender systems(CARS),this paper extends the advantages of graph convolutions to CARS.This paper first organizes the exogenous data as a unified attributed user-item bipartite graph,where context features are modeled as features of corresponding edges in the graph.Then,this paper proposes an end-to-end framework,which utilizes graph convolutional layers and factorization-machine-like decoder to model high-order collaborative and feature interaction patterns,respectively.Experiments on real-world datasets show that the proposed method achieves significant performance improvements compared to existing methods,and is expected to alleviate the cold start problem of items.Third,optimization objectives for graph-based recommendation:To solve the problems of inconformity with recommendation goal,amplification of popularity bias,and low training efficiency in existing optimization objectives,this paper redefines the task of recommendation from the perspective of noise contrastive estimation,resulting in the sampled softmax loss(SSM).Then this paper theoretically reveal three advantages of SSM:(1)alleviating popularity bias,(2)mining hard negative samples,and(3)maximizing DCG ranking metric.In addition,this paper also recognizes the potential shortcoming of SSM in learning the magnitudes of representation vectors,and provides mathematical proof that message passing schemes in graph convolution networks can adjust representation magnitude according to node degree,which naturally compensates for the shortcoming of SSM.Extensive experiments on benchmark datasets justify our analyses,demonstrating the superiority of training graph-based models using SSM for both normal and long-tail recommendation tasks.