Network representation learning (NRL) aims to encode the nodes and edges in complex networks into low dimensional vector space while preserving the inherent structure of the graphs. To learn better representations, three aspects of NRL, i.e., structural information in homogeneous graphs, relational information in heterogeneous graphs, and NRL to build side information for certain applications, e.g. recommendation systems, are extensively studied in the literature. However, sometimes two nodes in complex networks are located far away from each other. It is difficult for the NRL based models to capture the pairwise relationship with long distances. Besides, it would suffer from memory explosion if we aggregate all the high-order neighbors’ information in graphs. Hence, we use path learnin...[ Read more ]

Network representation learning (NRL) aims to encode the nodes and edges in complex networks into low dimensional vector space while preserving the inherent structure of the graphs. To learn better representations, three aspects of NRL, i.e., structural information in homogeneous graphs, relational information in heterogeneous graphs, and NRL to build side information for certain applications, e.g. recommendation systems, are extensively studied in the literature. However, sometimes two nodes in complex networks are located far away from each other. It is difficult for the NRL based models to capture the pairwise relationship with long distances. Besides, it would suffer from memory explosion if we aggregate all the high-order neighbors’ information in graphs. Hence, we use path learning to select high-order neighbors to enrich the information for the target nodes. Recently, path-based models have exhibited their power in many machine learning tasks. Inspired by the analysis of paths in both academia and industry, we propose path learning in complex networks in this thesis to address the three aspects of NRL in different ways.

In detail, we propose reinforce2vec, a biased random walk based approach for network embedding on homogeneous graphs. Our model uses a non-Markovian process to fully use the history of a random walk path. Second, to capture the inductive bias for learning PathSim based similarity scores, we propose NeuPath to identify a fixed number of path instances that can best infer the target meta-path(s) in heterogeneous information networks. Third, we study how the structural social information learned by path learning could be applied as the side information for the news recommendation task. We design an MCTS based method to explore high-order friends for the target user. A personalized hierarchical attention network is proposed for news recommendation on decentralized platforms. Besides, we construct a variant model for more industry-driven applications. Extensive experiments demonstrate the effectiveness and efficiency of the path learning in various complex networks.