Skip to main content

Table 3 Detailed results for five-fold cross-validation experiments, interpolating PageRank and Terrier scores (expansion of 20 related articles).

From: PageRank without hyperlinks: Reranking with PubMed related article networks for biomedical text retrieval

Tuning on MAP20
Fold training (λ = 0.7) testing (λ = 0.7) baseline
1 0.390 ± 0.275 0.514 ± 0.298 0.461 ± 0.304
2 0.447 ± 0.281 0.294 ± 0.260 0.264 ± 0.233
3 0.403 ± 0.273 0.473 ± 0.328 0.465 ± 0.313
4 0.439 ± 0.292 0.325 ± 0.225 0.277 ± 0.203
5 0.400 ± 0.284 0.477 ± 0.276 0.472 ± 0.300
Tuning on MAP40
Fold training (λ = 0.7) testing (λ = 0.7) baseline
1 0.513 ± 0.346 0.668 ± 0.344 0.636 ± 0.338
2 0.590 ± 0.343 0.365 ± 0.322 0.338 ± 0.290
3 0.539 ± 0.350 0.567 ± 0.356 0.560 ± 0.349
4 0.557 ± 0.343 0.495 ± 0.382 0.461 ± 0.382
5 0.400 ± 0.284 0.629 ± 0.317 0.627 ± 0.327
Tuning on P20
Fold training (λ = 0.6) testing (λ = 0.6) baseline
1 0.355 ± 0.330 0.520 ± 0.375 0.475 ± 0.366
2 0.421 ± 0.353 0.265 ± 0.277 0.250 ± 0.247
3 0.411 ± 0.349 0.289 ± 0.310 0.272 ± 0.299
4 0.379 ± 0.330 0.425 ± 0.404 0.400 ± 0.406
5 0.377 ± 0.348 0.435 ± 0.329 0.425 ± 0.338
  1. Mean and standard deviation for each fold are shown; for reference, baseline results on the test topics are provided. Note that optimized effectiveness metrics show consistent improvements over baseline.