Towards Co-Evolution of Random-Walk-Based Embedding and Label Propagation for Node Classification

We present a looped framework consisting of a task-specific node embedding module (XLAE) and a label propagation module (LP), to meet the challenges of node classification tasks. XLAE incorporates intermediate labels into node representations towards capturing local context information of both features and labels by running a Skip-gram model; LP leverages task-specific node embeddings to improve pairwise node proximity metric towards more accurate approximation of actual node labels. By this means of mutual promotion, XLAE and LP co-evolve in the loop and iteratively improve predictive accuracy. Our algorithm outperforms node embedding baselines on node classification over small subsets of labeled instances in both graph benchmark and social network datasets. With less parameters in the model and higher computational efficiency preserved, XLAE-LP closely matches or outperforms the state-of-the-art graph convolutional neural network approach on node classification.