Structure Discovery in fMRI Time-Series via Dynamic Graph Representation Learning

Structured health data often arrive as multivariate time series whose relational structure is unknown, subject-specific, and time-varying. Resting-state fMRI is a salient example: regional BOLD signals provide high-dimensional physiological measurements, but standard functional-connectivity pipelines usually impose a static Pearson-correlation graph. We propose Unsupervised Dynamic Graph Structure Learning (UDGSL), an unsupervised framework that infers a sequence of adjacency matrices directly from windowed fMRI time series. A temporal learner estimates region-region similarities for each window; a GCN encoder and decoder use the learned graph to reconstruct regional signals, yielding dynamic graph representations without labels or predefined connectivity. We evaluate UDGSL on HCP S1200 resting-state fMRI across AAL1 and Brainnetome parcellations. Compared with the best non-UDGSL MLP/LSTM baseline, UDGSL reduces reconstruction MSE by 24-32\%; relative to Pearson-correlation graphs, learned graphs reduce forecasting MSE by 26-28\% and improve subject-level classification accuracy by 5-7 absolute points. These results suggest that unsupervised graph structure learning can recover useful latent organization in complex physiological time series. We interpret the learned graphs as predictive, data-driven relational representations rather than causal or clinical connectivity estimates, and discuss extensions to signed and static-dynamic graph models.