GuidedWalk

General-purpose embeddings do not guarantee to produce the best representation for the target tasks. In the graph domain, random walk-based graph embedding methods like DeepWalk and Node2Vec are widely used methods that generate general-purpose embedding. These methods, however, can not achieve high accuracy in different tasks, including node classification. In contrast, semi-supervised methods such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), and their extensions achieve state-of-the-art performance in this task as they use few labels for learning the most appropriate embedding. These methods, however, depend on node features and cannot achieve high performance in the absence of node features or low dimension features. In this paper, we propose GuidedWalk, a semi-supervised random walk-based graph embedding method that can outperform other random walk-based competitors as well as GNNs in the semi-supervised setting on graphs without node features. The proposed model works based on exploring graph paths with more emphasis on the paths that connect nodes of the same class. We show that the neural processing core of DeepWalk and Node2Vec propagates latent features across sampled paths. Therefore, our selected paths increase the chances of propagating the right latent features, which is appropriate for the node classification task. Our experiments on Cora, Pubmed, Twitch DE, and Facebook datasets, show 0.53%, 0.78%, 5.07%, and 7.13% improvement in node classification accuracy compared to the state-of-the-art techniques in the field.