Table 1. Edge prediction with BTL as a benchmark.
During 50 independent trials of fivefold cross-validation (250 total folds per network), columns show the percentages of instances in which SpringRank Eq. 3 and regularized SpringRank Eq. 5 with α = 2 produced probabilistic predictions with equal or higher accuracy than BTL. Distributions of accuracy improvements are shown in Fig. 3. Center columns show accuracy σa, and right columns show σL (Materials and Methods). Italics indicate where BTL outperformed SpringRank for more than 50% of tests. NCAA Basketball data sets were analyzed 1 year at a time.
| Data set | Type | % Trials higher σa versus BTL | % Trials higher σL versus BTL | ||
| SpringRank | +Regularization | SpringRank | +Regularization | ||
| Computer science (3) | Faculty hiring | 100.0 | 97.2 | 100.0 | 99.6 |
| Alakāpuram (2) | Social support | 99.2* | 99.6 | 100.0 | 100.0 |
| Synthetic β = 5 | Synthetic | 98.4 | 63.2 | 76.4 | 46.4 |
| History (3) | Faculty hiring | 97.6* | 96.8 | 98.8 | 98.8 |
| NCAA Women (1998–2017) (39) | Basketball | 94.4* | 87.0 | 69.1 | 51.0 |
| Tenpaṭṭi (2) | Social support | 88.8 | 93.6 | 100.0 | 100.0 |
| Synthetic β = 1 | Synthetic | 83.2 | 65.2 | 98.4 | 98.4 |
| NCAA Men (1985–2017) (39) | Basketball | 76.0* | 62.3 | 68.5 | 52.4 |
| Parakeet G1 (5) | Animal dominance | 71.2* | 56.8 | 41.2 | 37.2 |
| Business (3) | Faculty hiring | 66.8* | 59.2 | 39.2 | 36.8 |
| Parakeet G2 (5) | Animal dominance | 62.0 | 51.6 | 47.6 | 47.2 |
*Tests that are shown in detail in Fig. 4.