ModTGCN: Faster, better text classification via modularity-aware GNNs

Traditional graph-based text classification models rely on local neighborhood aggregation, often ignoring global community structure despite semantic document graphs exhibiting strong class-consistent clustering. This oversight can blur class boundaries and cause over-smoothing. To address this, researchers introduce ModTGCN, a modularity-aware graph neural network that jointly optimizes cross-entropy loss with a modularity-based auxiliary objective. The modularity term is computed on a document-document similarity graph derived from transformer embeddings (pretrained or fine-tuned). By explicitly promoting class-coherent document communities, ModTGCN preserves discriminative representations while leveraging global structure.

Key architectural innovations drive performance and efficiency. The model decouples the original heterogeneous TextGCN graph into separate document-word and word-word components, resulting in 2x-10x faster training compared to prior approaches. Researchers also explore graph construction strategies, label-aware edge reweighting, and supervision choices for modularity optimization. Evaluated on five standard benchmarks, ModTGCN achieves consistent accuracy gains, with the largest improvements on complex, low homophily datasets such as Ohsumed and 20NG. The work, accepted at PAKDD 2026, offers a practical path to more accurate and scalable text classification without sacrificing speed.