Network Representation Learning Based Recommender Systems

The explosive growth of online content and services has provided overwhelming choices for users,such as news,movies,books,restaurants and music.Recommender systems(RS)intend to address the information explosion by finding a small set of items for users to meet their personalized interests.The key idea of RS is modeling user-related and item-related information(e.g.,user history,item attributes,contexts),therefore,a practical recommendation algorithm should be able to incorporate various side information.In this thesis,we focus on a special type of side information,i.e.,network-structured information.For example,there may be an online social network among users and a knowledge graph(KG)among items,and the useritem interaction can also be treated as an interaction graph.Given the rich network-structured information,it is a great challenge how to effectively utilize the high-dimensional data for RS.Recently,network representation learning(NRL)is becoming a new research direction in the field of machine learning.NRL 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 networks in RS,it is a promising approach to combine NRL with RS.Using NRL to process the network data in RS can improve the capacity of RS and boost the precision as well as user satisfaction of recommended results,thereby enhancing the overall economic efficiency.In this thesis,we investigate network representation learning based recommender systems.The major contributions of this thesis are as follows:First,we study network representation learning methods applied to the interaction graph in recommender systems.The explicit or implicit feedback between users and items comprises an interaction graph in RS.Therefore,we propose designing RS from the perspective of NRL.We present GraphGAN,a framework unifying generative and discriminative NRL methods,in which the generator and discriminator play a game-theoretical minimax game.Specifically,for a given vertex,the generator fits its underlying true connectivity distribution over all other vertices and produces“fake”neighbors to fool the discriminator,while the discriminator detects whether the sampled vertex is from ground truth or generated by the generator.The learned embeddings can be used to characterize users/items and applied to RS.Second,we study social-network-aware recommender systems.According to the homophily assumption,two closely related users in a social network may also share similar preferences in RS.Therefore,we present two methods that leverage the social network to assist with RS:(1)Embedding-based method first uses NRL to map each user in the social network to a low-dimensional space,then applies the learned embeddings to RS task.Specifically,we propose SHINE model that uses auto-encoder to mine users' social relationship for Weibo celebrity recommendation.(2)Structure-based method leverages the structure of the social network more directly.Specifically,we study the influence of the social network on user participation of Weibo online voting.We design a joint matrix factorization model JTS-MF,which incorporates the users' social relationship and group information into RS.The experiment results consistently demonstrate that a social network is essential to improve the performance of RS.Third,we study knowledge-graph-aware recommender systems.A KG enriches the description of items and strengthens the inter-item relatedness,which is valuable for RS.Similarly,we present two types of methods that leverage the KG for RS:(1)Embedding-based methods use knowledge graph embedding(KGE)approaches to learn entity and relation representation vectors for the subsequent RS.Based on the training order of the KGE task and the RS task,embedding-based methods can be further categorized into one-by-one learning and alternate learning.We present two models accordingly: DKN uses a knowledge-aware convolutional neural network and an attention network to learn representations of news titles and aggregate users' historical interests,respectively;MKR deploys a multi-task learning framework leveraging the KGE task to assist with the RS task.(2)Structure-based methods use breadth-first-search to obtain multi-hop neighbors of an entity in the KG.Structure-based methods can be categorized into outward propagation and inward aggregation: We present RippleNet,a representative of outward propagation that propagates user preferences on the KG to discover their hierarchical potential interests;We also present KGCN,a representative of inward aggregation that learns the representation of an entity by aggregating information from its multi-hop neighbors to mine users' potential preferences.The experiment results show that utilizing high-order structural information of the KG can greatly benefit RS.In addition,embedding-based methods have high flexibility while structure-based methods have high explainability.