The increasing complexity of graph neural networks (GNNs) necessitates enhanced interpretability to establish trust in real-world deployments. While subgraph-based explanations improve human understanding, current methods face critical limitations: stochastic subgraph sampling induces bias in instance-level interpretations, and labor-intensive manual comparisons hinder efficient analysis. To address these challenges, we propose a classification-driven framework that identifies class-consistent subgraph patterns through graph similarity computation, reducing sampling randomness while providing model insights at the categorical level. Additionally, we develop GEVis, a visual analytics system that supports multi-faceted evaluation of explanation algorithms and model decisions through interactive exploration, enabling efficient discovery of hidden data relationships without requiring specialized expertise. Experimental validation across synthetic, molecular, and textual datasets demonstrates the framework’s feasibility, with performance variations across data types aligning with structural complexity differences. Qualitative case studies and quantitative efficiency metrics further confirm GEVis’s practical utility in interpretability analysis.