Adaptive Graph Sparsification for Efficient Graph Neural Networks

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This preprint presents an experimental study of graph sparsification for graph neural networks. The study evaluates multiple edge-selection strategies, including random, degree-based, structural, feature-based, and multi-criteria sparsification, across the Cora, CiteSeer, PubMed, Chameleon, and Squirrel datasets. The experiments examine the relationship between edge retention and predictive performance using test accuracy and macro-F1, together with training and computational time. Pairwise ablation experiments further evaluate the contributions of structural similarity, node-feature similarity, and local connectivity components within the proposed multi-criteria framework. The study also analyzes graph characteristics and computational performance across datasets with different structural and homophily properties. Results are reported across multiple retention ratios and random seeds. The uploaded document contains the research methodology, experimental results, ablation analysis, discussion, limitations, and references.