PageRank-Guided Random Walks for Improved Graph Representations

Graph embedding efficiently maps complex graph features within a low-dimensional vector space, facilitating a variety of downstream network prediction tasks. Recent advancements often leverage random walk-based methods due to their proficiency in capturing the structural information of graphs. However, these methods tend to exhibit an inherent bias toward high-centrality nodes, which frequently serve as pivotal bridges in network topology. Such biases disproportionately influence the learning process and can skew embeddings. This leads to inaccurate predictions, particularly for nodes with sparse connections. This paper proposes a novel approach that utilizes the PageRank algorithm to quantify node importance and identify imbalances. By integrating these insights, we aim to mitigate centrality bias in graph embeddings, thus enhancing the representational fairness and predictive accuracy of the model.