Towards Local Information Enhanced Representation Learning On Networks

Network representation learning is an effective approach for mining and utilizing large-scale,diverse,and complex data in real-world scenarios.It is like a bridge con-necting the input network and downstream tasks and it aims to represent nodes as low-dimensional dense real value vectors according to the structure information and attribute information in the input network.The learned representations can be used as features for downstream tasks.Due to the prevalent network structure data in the real world,network representation learning has received increasing attention from researchers and become an active research area.Although existing methods combine structure and node attribute information for representation learning,it is not enough for sufficiently mining the local information of nodes and further takes its toll on downstream tasks.In this article,to im-prove the expressive ability of nodes learned by network representation learning model,we start with making further efforts on mining the local feature information of the nodes and exploring the local information enhanced representation learning methods.Specif-ically,we study two problems:learning relationship-preserving heterogeneous network representations and the representation learning of tail nodes on networks.Through local information enhancement,we can express the information of a node comprehensively and improve the quality of the representation.First,we leverage metagraphs to mining the relationship of nodes and their sur-rendering graph structure,enhancing nodes' awareness of local semantic and structure information.And in our model,metagraphs actively participate in the learning pro-cess by mapping themselves to the same embedding space as the nodes do.Moreover,metagraphs guide the learning through both first-and second-order constraints on node embeddings,to model not only latent relationships between a pair of nodes,but also individual preferences of each node.Second,considering the long-tailed distribution on network,we employ MAML for the meta-learning of tail node representation,which is capable of enhancing the nodes' awareness of local information.So that the few shot regression model learns to learn the parameters for the tail nodes at different localities and refines the tail node representations based on local information in the reconstruction progress.Results of extensive experiments on several public datasets show that our approach significantly outperforms a suite of state-of-the-art baselines in downstream tasks.