MetaGNN Based Heterogeneous Information Networks Representation Learning

Heterogeneous Information Networks(HINs)are ubiquitous in many real-world scenarios,such as social networks,academic citation networks,movies rating networks and etc.It is important yet challenging to mine valuable information from these networks,and the main challenge is how to build the proper information networks representation.The classical information networks representation models view information networks as the graph models which consist of nodes and edges,and manually select nodes' features as their representation.However,these models highly rely on experts' prior knowledge and consume much time.To address these problems,many information networks representation learning algorithms are proposed,which aims to automatically learn nodes' representation.Graph neural networks(GNNs)based information networks representation learning algorithms have attracted considerable interests because of their good performance.They assume that adjacent nodes have the similar attributes(e.g.nodes' labels)and obtains target nodes' low-dimension vector representation by aggregating neighbor nodes features.However,most existing GNNs based methods have some problems.First,they only focus on the homogeneous networks with one type of nodes and edges,which cannot generalize to HINs because various types of nodes have different degree influence on target nodes.Second,manually designing nodes aggregating strategies highly rely on experts' knowledge.To address these problems,we propose MetaGNN,a hierarchical networks representation learning algorithm which consists of Deep Q-network(DQN)component and GNNs component.The DQN component automatically learns to aggregate the various types of neighbor nodes of target nodes from GNNs component.GNNs component learns the nodes vector representation according to DQN's aggregating strategies and feedbacks the new target nodes and downstream task related evaluation metric to DQN component as new input and reward respectively.The proposed method has the following three advantages: First,the proposed model is an end-to-end model which can directly output downstream task related evaluation metric.Second,it can automatically learn the aggregating strategies for various types of neighbor nodes without any experts' prior knowledge.Third,it can effectively generalize to new nodes.The model has been validated on three real-world large-scale datasets on both inductive and transductive nodes multi-classification tasks and obtained promising results comparing with four classical information networks representation learning algorithms.