Graph Convolutional Networks (GCNs) have become a cornerstone of graph-based machine learning, but their performance degrades with increased depth due to the over-smoothing phenomenon—where node representations become indistinguishable. This paper presents a focused comparative study of regularization techniques specifically designed to address over-smoothing. We evaluated four representative methods: DropNode, DropEdge, PairNorm, and Jumping Knowledge Networks. Each is analyzed in terms of its effectiveness in preserving node-level distinctions, scalability across network depths, and overall classification performance. Experiments on standard benchmarks demonstrate the trade-offs and practical implications of each technique. Our findings provide actionable information for selecting appropriate regularization strategies and highlight promising directions to improve GCN robustness at scale.