LLaCE: Locally Linear Contrastive Embedding

Node embedding is one of the most widely adopted techniques in numerous graph analysis tasks, such as node classification. Methods for node embedding can be broadly classified into three categories: proximity matrix factorization approaches, sampling methods, and deep learning strategies. Among the deep learning strategies, graph contrastive learning has attracted significant interest. Yet, it has been observed that existing graph contrastive learning approaches do not adequately preserve the local topological structure of the original graphs, particularly when neighboring nodes belong to disparate categories. To address this challenge, this paper introduces a novel node embedding approach named Locally Linear Contrastive Embedding (LLaCE). LLaCE is designed to maintain the intrinsic geometric structure of graph data by utilizing locally linear formulation, thereby ensuring that the local topological characteristics are accurately reflected in the embedding space. Experimental results on one synthetic dataset and five real-world datasets validate the effectiveness of our proposed method.