Effective representation learning for graph-structured data with adversarial learning

Abstract

Graph-structured data is widely existed in real-world applications such as social networks, paper citation networks and protein-protein interaction networks. It encodes very rich information about data entities in graph structures through the complicated connections among them. How to extract such abundant information is an important and challenging problem which has attracted a great amount of attention from both academia and industry. Traditional methods rely on hand-engineered features which are both ineffective and inefficient. In recent years, representation learning emerges as the most promising way for modeling graph-structured data, which aims to learn low-dimensional vectors for nodes in the graph. The learned node representations can further be utilized to facilitate downstream learning tasks such as network analysis (e.g., node classification and link prediction), recommendation, and fraud detection. This technique is called graph representation learning or network embedding in the literature. In this thesis, we aim to learn effective node representations for both plain networks and attributed networks with the assistance of adversarial learning including adversarial learning principle based on generative adversarial networks (GANs) and adversarial training methods from adversarial machine learning. For plain network embedding, we design regularization methods and sampling method to enhance embedding learning. Specifically, we first propose a global regularization method for deep embedding models via GANs, of which a prior distribution is imposed on embedding vectors to help alleviate overfitting. Then, we introduce a succinct and effective local regularization method, namely adversarial training, for negative sampling based embedding models such as DeepWalk, LINE and node2vec, which can improve both model robustness and generalization performance. Furthermore, we propose an adversarial ranking network embedding model to preserve node similarity rankings in representations, which unifies a triplet sampling phase and an embedding learning phase with the framework of GANs. It can encourage the generation of more difficult and relevant negative nodes for given positive target-context node pairs to improve representation learning. For attributed network embedding, we focus on a challenging cross-network learning problem that aims to transfer the label information from an attributed source network to an attributed target network. Specifically, we propose a novel network transfer learning framework AdaGCN via adversarial domain adaptation and graph convolution, which enables the learning of both class discriminative and domain invariant node representations and thus facilitates cross-network node classification. Extensive empirical evaluations on benchmark datasets demonstrate the effectiveness of the proposed methods.