Learning Network-Invariant and Label-Discriminative Representations for Cross-Network Node Classification

Networks are ubiquitous data structures in the real world. The accurate and efficient analysis of networks is critical to realizing many intelligent network-based services. However, most existing network analysis methods are developed for single networks and require a lot of labeled data, which is costly and time-consuming to acquire. Transfer learning has been widely accepted as an effective paradigm for tackling this problem by reusing the model trained on a supervised task. However, transfer learning on the non-euclidean network data has been investigated by no more than a few studies. To realize accurate node classification based on the knowledge learned from the labeled source network, this paper proposes to learn network-invariant and label-discriminative representations based on graph embedding and linear discriminant analysis. Specifically, we embed the source and target networks into adjacent vector spaces based on the graph attention network by minimizing the Sinkhorn distributional distances between their embeddings. To obtain label-discriminative features for learning better classification models, we then utilize a transferable linear discriminative analysis method to project the embeddings into label-discriminative subspaces. In the end, a support vector machine model trained on the labeled source network is utilized to classify the target nodes. Experiments on two pairs of networks illustrate that our method achieves the best performance and evaluates the effectiveness of the proposed modules.