GW Geodesic Mixup

Integrate this into the minibatch data loader for graph classification, molecular prediction, or attributed-network learning. Select two graphs X0 and X1, compute a node coupling Pi using entropic GW or a cheaper feature-based Sinkhorn cost, and sample t from Beta(alpha,alpha), starting with alpha=0.4. Construct a mixed graph with the node set of X0. For each source edge or relational pair (i,j), choose target indices k=argmax Pi[i,:] and l=argmax Pi[j,:] for a deterministic MVP, then set the mixed attribute to Geo_t(omega_0[i,j],omega_1[k,l]). For Euclidean attributes use (1-t)omega_0[i,j]+t omega_1[k,l]; for vector node features use the same correspondence and interpolation. A stochastic version samples (k,l) with probability proportional to Pi[i,k]Pi[j,l], while a soft version averages the resulting attributes. Train on the soft label (1-t)y0+t y1, or use endpoint labels if the task is not linearly compositional. The paper supplies the pointwise constant-speed geodesic construction and requires Z to be geodesic; Pi, discrete matching, and finite-size approximations are estimated numerically. First compare a small GIN against standard graph mixup, permutation augmentation, and no augmentation on synthetic permuted graphs, then MUTAG or ZINC. Measure accuracy, calibration, robustness to node deletion, and whether estimated distances to endpoints scale approximately as t and 1-t. The expected signal is improved interpolation-region accuracy and better validation performance with fewer labeled examples.