<p>Functional magnetic resonance imaging (fMRI) is widely used to characterize functional brain organization through graph-based connectome analysis. However, functional connectivity networks constructed using Pearson correlation are typically dense and noisy, which compromises the stability of network topology measures, limits interpretability, and degrades the performance of downstream tasks such as graph-based analysis. Consequently, effective brain graph sparsification has become a critical step for improving the reliability and modeling efficiency of fMRI-based network analysis. Although a growing number of sparsification methods have been proposed in recent years, existing approaches remain fragmented and a structured methodological synthesis of this area is still lacking. To address this gap, we provide a methodological review of fMRI-based brain graph sparsification techniques. According to whether edge selection is governed by intrinsic graph properties or informed by external supervision signals, we provide a taxonomy-driven methodological review of fMRI-based brain graph sparsification. We organize existing methods into topology-guided and supervision-guided paradigms, clarify their underlying principles and trade-offs, and propose an evaluation framework and reporting considerations to enhance transparency and reproducibility. We further outline key challenges and future directions toward robust and biologically meaningful connectome modeling.</p>