Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction

Version 0.1.0 of the data companion to “Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction.” This release contains approved derived resting-state fMRI functional connectivity data and reproducibility metadata for benchmarking the external generalization of symmetric positive-definite matrix learning methods. Each connectome is represented as a 100 × 100 Schaefer-parcellation correlation matrix obtained through regularized covariance estimation and numerical stabilization. The archive contains approved derived connectomes and release-safe metadata for CamCAN, ABIDE, and COBRE, together with dataset inventories, data dictionaries, provenance records, validation reports, configuration snapshots, and restricted reconstruction documentation. It does not contain raw or preprocessed MRI, ROI time series, or original or hashed participant identifiers. The associated benchmark evaluates chronological-age regression under subject-grouped cross-validation and leave-one-dataset-out external validation, using tangent-space Ridge, SPDNet, vectorized-correlation Ridge, and dummy baselines. Dataset-specific licenses, required citations, acknowledgements, and reuse conditions are provided in LICENSES.md. No single record-level license overrides those dataset-specific terms.