We study the impact of data partitioning strategies, feature families, and graph topology on supervised link prediction over large-scale academic collaboration networks. In contrast to standard node-partitioned evaluation protocols, we analyze a less studied but widely used practice in applied graph mining: edge-ordered training, where positive and negative links are generated from a single evolving graph and then ordered or shuffled before splitting. Using two real collaboration graphs (UNAM and Stanford), we evaluate multiple supervised pipelines combining semantic Feature Proximity (FP), Local Structural Proximity (LSP), and Global Structural Proximity (GSP), trained under both edge-ordered and node-partitioned regimes. Eight controlled experimental pairs are constructed to isolate the effects of (i) feature type (semantic vs. topological), (ii) feature cardinality, (iii) graph origin (UNAM vs. Stanford), and (iv) data partitioning protocol (edges vs. nodes). Models include classical machine-learning classifiers as well as stacked meta-models trained over base predictors. Performance is assessed using AUC-ROC, AUC-PR, F1, and Accuracy across repeated runs, and statistical significance is evaluated using paired Wilcoxon signed-rank tests. Our results indicate that node-partitioned training often yields higher AUC-ROC and F1 when semantic and mixed FP+LSP+GSP features are used, while edge-ordered training tends to produce higher scores for purely topological features, a pattern compatible with information leakage induced by shared neighborhoods. Furthermore, meta-models frequently achieve nearperfect scores under edge-ordered settings but degrade substantially under node-partitioned splits, suggesting sensitivity to leakage-driven signals rather than purely generalizable structure. Keywords: link prediction, scientific collaboration networks, graph-based machine learning, edge-ordered, node-partitioned