Adaptive anchor-based attention networks for large-scale sparse bipartite graph embedding

doi:10.1016/j.knosys.2025.114242

Adaptive anchor-based attention networks for large-scale sparse bipartite graph embedding

2025-08-09

0

PRE

AI

Renata Borovica‐Gajic

DOI：10.1016/j.knosys.2025.114242

Original

Original request for help

Save

Abstract

En 中文

Bipartite graph embedding aims to map each node into compact, low-dimensional vectors that preserve the intrinsic properties of the graph. The effectiveness of these embeddings is crucial for capturing structural information and node relationships, which directly impacts the performance of downstream applications such as recommender systems and bioinformatics. However, due to the inherent sparsity and large scale of many real-world bipartite graphs, existing methods often suffer from missing contextual information and excessive feature smoothing. In this paper, we take the first step toward systematically addressing the challenges of embedding large and sparse bipartite graphs. To this end, we propose Adaptive Anchor-based Graph Attention Networks (A 2GAT), a novel framework that integrates entropy regularization to ensure a balanced distribution of attention weights, preserve feature distinctiveness, and mitigate over-smoothing. In addition, we design an adaptive anchor node generation mechanism and introduce a fully connected attention (FCA) module that dynamically adjusts interaction weights, effectively addressing sparse connectivity and enhancing representation learning in low-density regions. Extensive experiments on eight benchmark datasets demonstrate the effectiveness and generalizability of our model across both recommendation and link prediction tasks.

Keywords:

bipartite graph embedding

attention mechanism

entropy regularization

over-smoothing

anchor nodes

Adaptive anchor-based attention networks for large-scale sparse bipartite graph embedding

Abstract

Journal

Organization