Graph Representation Learning-Based Recommender Systems

Personalized recommendation has been applied to many online services such as E-commerce and adverting.It facilitates users to discover a small set of relevant items,which meet their personalized interests,from many choices.Nowadays,diverse kinds of auxiliary information on users and items become increasingly available in online platforms,such as user demographics,social relations,and item knowledge.More recent evidence suggests that incorporating such auxiliary data with collaborative filtering can better capture the underlying and complex user-item relationships,and further achieve higher recommendation quality.In this dissertation,we focus on auxiliary data with graph structure,such as social networks and Knowledge Graphs(KG).For example,we can improve recommendation performance by mining social relationships between users,and by using knowledge graphs to enhance the semantics of recommended items.Network representation learning aims to represent each vertex in a network(graph)as a low-dimensional vector while still preserving its structural information.Due to the existence of massive graph data in recommender systems,it is a promising approach to combine network representation learning with recommendation.Applying the learned graph features to recommender systems will effectively enhance the learning ability of the recommender systems and improve the accuracy and user satisfaction of the recommender systems.For network representation learning and its application in recommendation systems,the major contributions of this dissertation are as follows:(1)Attention-based Adversarial Autoencoder for Multi-scale Network Embedding.Existing Network representation methods usually adopt a “one-size-fits-all” approach when concerning multi-scale structure information,such as first-and second-order proximity of nodes,ignoring the fact that different scales play different roles in embedding learning.We propose an Attention-based Adversarial Autoencoder Network Embedding(AAANE)framework,which promotes the collaboration of different scales and lets them vote for robust representations.(2)Context-Dependent Propagating-based Video Recommendation in Multimodal Heterogeneous Information Networks.Conventional approaches to video recommendation primarily focus on exploiting content features or simple user-video interactions to model the users' preferences.However,these methods fail to model complex video context interdependency,which is obscure/hidden in heterogeneous auxiliary data.In this dissertation,we propose a Context-Dependent Propagating Recommendation network(CDPRec)to obtain accurate video embedding and capture global context cues among videos in Heterogeneous Information Networks(HINs).The CDPRec can iteratively propagate the contexts of a video along links in a graph structured HIN and explore multiple types of dependencies among the surrounding video nodes.Besides,Graph structure is also highly vulnerable to small but intentional perturbation attacks on the input video features,since the perturbations from connected nodes can aggregate the impact on a target video node in the graph.To capture the heterogeneous relation and enhance the robustness of Graph Neural Network(GNN)for recommendation,we propose a new optimisation framework,namely Adversarial Heterogeneous Graph Neural Network for video RECommendation(AHGNNRec).AHGNNRec apply an Adversarial Training(AT)method to optimise the embedding propagation layers.AT can be seen as playing a minimax game: the generated adversarial fake nodes from clean nodes with perturbations maximally attacking the recommendation objective,and then we learn over these adversarial nodes by minimising impact of the additional adversarial regularisation term.(3)Knowledge Graph Enhanced Neural Collaborative Filtering.Existing KG based recommendation methods mainly rely on handcrafted meta-path features or simple triple-level entity embedding,which cannot automatically capture entities' long-term relational dependencies for recommendations.In this dissertation,a two-channel neural interaction method named Knowledge Graph Enhanced Neural Collaborative Filtering is proposed,which leverages both long-term relational dependency KG context and user-item interaction for recommendations.For the KG context interaction learning,we propose a Residual Recurrent Network(RRN)to construct context-based path embedding.RRN incorporates residual learning into traditional recurrent neural networks(RNN)to efficiently encode the long-term relational dependencies of KG.