2026 Volume 34 Pages 818-829
Random walk algorithms are widely used for graph-based search and exploration but are often limited by their reliance on local node information. This study introduces EmbedRW (Embedded Random Walk), a novel random walk algorithm that leverages node embedding vectors to inform transition probabilities. Unlike traditional methods that rely solely on structural connectivity, EmbedRW calculates transition probabilities based on the distance in the latent space between the neighbor nodes and the target node, thereby incorporating global semantic proximity into the local traversal process. By integrating node embeddings—such as those generated by Node2vec or DeepWalk—into the agent's decision-making, EmbedRW aims to improve the efficiency of locating target nodes. We evaluate EmbedRW across diverse network topologies and varying graph densities. The experimental results demonstrate that EmbedRW empirically reduces average hitting times by up to 90% compared to a variety of conventional random walk methods, such as Simple Random Walk (SRW) and Metropolis-Hastings Random Walk (MH). The proposed method is particularly effective in dense graphs where multiple paths to the target exist. Furthermore, our analysis shows that EmbedRW remains effective and outperforms random search strategies even when embedding vectors are available for only a subset of nodes. However, its impact on cover time — the expected time to visit all nodes — is limited. This limitation arises because the embedding-based bias facilitates rapid convergence toward specific targets, which inherently conflicts with the diffusive behavior required for exhaustive graph exploration.
