PhaaNe's picture
KGGuard+W checkpoints, scores, data, caches and references (code commit 3616af2) (part 10)
2f94d1e verified
|
Raw History Blame Contribute Delete
30.3 kB

Robustness of the final method (E6)

Design split: mean over 3 seeds. Fresh splits: mean over splits, and how many splits the variant is better than the default.

1. Hyper-parameter sensitivity (test AP; dAP vs default; CV dAP = same change on the cross-fit estimate)

variant dbpedia_accuracy AP dbpedia_accuracy dAP dbpedia_accuracy CV dAP dbpedia_accuracy ms dAP dbpedia_accuracy ms better factbench AP factbench dAP factbench CV dAP factbench ms dAP factbench ms better
default (final configuration) 0.4697 0 0 0 0/5 0.9285 0 0 0 0/5
calibration bins = 5 0.4657 -0.004 -0.0015 -0.0045 1/5 0.9285 0.0001 -0.0151 -0.0074 1/5
calibration bins = 20 0.4741 0.0043 -0.0013 -0.0023 1/5 0.9311 0.0026 -0.0022 0.0019 5/5
smoothing kappa = 5 0.4709 0.0011 -0.0015 -0.0004 1/5 0.9289 0.0004 0.0004 -0.0002 1/5
smoothing kappa = 50 0.4699 0.0002 0.001 0.001 4/5 0.9285 0 -0.0001 -0.0002 2/5
cross-fit folds k = 3 0.4718 0.0021 0.0038 0.0018 3/5 0.9283 -0.0002 -0.0049 -0.0002 2/5
cross-fit folds k = 10 0.4666 -0.0031 0.0006 0.0024 4/5 0.9286 0.0001 0.0003 0.0005 3/5
fold assignment seed = 1 0.4746 0.0048 0.0096 -0.0046 1/5 0.9276 -0.0009 -0.0004 0.0001 2/5
fold assignment seed = 2 0.4635 -0.0062 0.0065 -0.0033 1/5 0.929 0.0006 -0.0046 -0.0003 2/5
in-sample prior as calibration reference (no out-of-fold) 0.4745 0.0048 0.002 0.002 3/5 0.9285 0 0.0003 0.0001 3/5
GAM learning_rate = 0.025 0.4638 -0.0059 -0.0039 -0.0035 0/5 0.9275 -0.001 -0.0012 -0.0012 1/5
GAM learning_rate = 0.1 0.4725 0.0028 0.0017 0.001 4/5 0.9285 0 -0.0007 -0.0001 3/5
GAM max_iter = 150 0.464 -0.0058 -0.0038 -0.0035 0/5 0.9275 -0.0009 -0.0012 -0.0013 1/5
GAM max_iter = 600 0.4728 0.0031 0.0017 0.001 4/5 0.9285 0 -0.0007 -0.0001 3/5
GAM max_leaf_nodes = 4 0.4672 -0.0025 0.0007 -0.0007 1/5 0.9276 -0.0009 -0.0003 0 3/5
GAM max_leaf_nodes = 16 0.4698 0.0001 -0 0 2/5 0.9285 0 0 -0 2/5
GAM min_samples_leaf = 20 0.4721 0.0023 0.0076 0.0003 2/5 0.9285 0 -0.0001 -0 0/5
GAM min_samples_leaf = 80 0.4699 0.0001 -0.0003 -0.0004 2/5 0.9285 0 0 0.0001 4/5
GAM l2_regularization = 0 0.4701 0.0004 0.0002 -0.0003 1/5 0.9291 0.0006 0 0.0002 4/5
GAM l2_regularization = 5 0.4684 -0.0013 -0.0009 -0.0005 2/5 0.9274 -0.0011 0 -0.0001 1/5
GAM without monotone constraints 0.4634 -0.0063 -0.0022 -0.0111 0/5 0.9429 0.0144 0.022 0.0225 5/5

2. Missing-evidence strategy (AP; 'X missing (natural)' = test facts whose source X is missing in the data)

('dbpedia_accuracy', 'design') ('dbpedia_accuracy', 'fresh splits') ('factbench', 'design') ('factbench', 'fresh splits')
('C missing (natural)', 'GBDT reference') 0.6301 0.5971 nan nan
('C missing (natural)', 'correct') 0.5726 0.572 nan nan
('C missing (natural)', 'mean') 0.5823 0.5742 nan nan
('C missing (natural)', 'native') 0.6002 0.5795 nan nan
('C missing (natural)', 'zero (proposed)') 0.5846 0.5755 nan nan
('S missing (natural)', 'GBDT reference') nan nan 0.9718 nan
('S missing (natural)', 'correct') nan nan 0.5355 nan
('S missing (natural)', 'mean') nan nan 0.6141 nan
('S missing (natural)', 'native') nan nan 0.9235 nan
('S missing (natural)', 'zero (proposed)') nan nan 0.6133 nan
('S_sup missing (natural)', 'GBDT reference') 0.6137 0.541 nan nan
('S_sup missing (natural)', 'correct') 0.5976 0.5357 nan nan
('S_sup missing (natural)', 'mean') 0.6161 0.5296 nan nan
('S_sup missing (natural)', 'native') 0.6191 0.5365 nan nan
('S_sup missing (natural)', 'zero (proposed)') 0.6214 0.5329 nan nan
('T missing (natural)', 'GBDT reference') nan 0.475 0.9164 0.9429
('T missing (natural)', 'correct') nan 0.5 0.7967 0.7476
('T missing (natural)', 'mean') nan 0.4513 0.8279 0.7965
('T missing (natural)', 'native') nan 0.4914 0.8514 0.8239
('T missing (natural)', 'zero (proposed)') nan 0.4509 0.8307 0.7953
('W missing (natural)', 'GBDT reference') nan nan 0.9956 1
('W missing (natural)', 'correct') nan nan 0.9986 0.9985
('W missing (natural)', 'mean') nan nan 0.9929 0.9882
('W missing (natural)', 'native') nan nan 0.9925 0.9836
('W missing (natural)', 'zero (proposed)') nan nan 0.9936 0.9882
('all', 'GBDT reference') 0.4943 0.5019 0.9814 0.9836
('all', 'correct') 0.4615 0.49 0.9039 0.894
('all', 'mean') 0.4672 0.4937 0.9288 0.9276
('all', 'native') 0.4755 0.497 0.9511 0.9526
('all', 'zero (proposed)') 0.4697 0.4945 0.9285 0.928

3. Test-time source dropout, frozen models (dAP vs the same model without dropout; design split and fresh splits pooled)

dbpedia_accuracy dAP factbench dAP dbpedia_accuracy AP factbench AP
('C', 0.1, 'correct') 0.0005 -0.0035 0.4798 0.8942
('C', 0.1, 'mean') 0.0003 -0.0012 0.484 0.9268
('C', 0.1, 'native') 0.0003 -0.0011 0.4893 0.9509
('C', 0.1, 'zero (proposed)') 0.0003 -0.0019 0.4855 0.9262
('C', 0.3, 'correct') -0.0003 -0.0058 0.479 0.8919
('C', 0.3, 'mean') 0.0003 -0.0029 0.484 0.9251
('C', 0.3, 'native') -0.0006 -0.0024 0.4884 0.9496
('C', 0.3, 'zero (proposed)') -0.0004 -0.0035 0.4848 0.9246
('C', 0.5, 'correct') -0.0011 -0.0091 0.4783 0.8886
('C', 0.5, 'mean') -0.0003 -0.0051 0.4835 0.9229
('C', 0.5, 'native') -0.0007 -0.0029 0.4883 0.9491
('C', 0.5, 'zero (proposed)') -0.0009 -0.0055 0.4843 0.9226
('C', 1.0, 'correct') -0.0009 -0.0158 0.4785 0.8818
('C', 1.0, 'mean') -0.0003 -0.0099 0.4834 0.9182
('C', 1.0, 'native') -0.0005 -0.0053 0.4884 0.9467
('C', 1.0, 'zero (proposed)') -0.0007 -0.01 0.4845 0.9182
('M', 0.1, 'correct') -0.0008 -0.0017 0.4785 0.896
('M', 0.1, 'mean') -0.0002 0.0001 0.4835 0.9282
('M', 0.1, 'native') -0.0014 0 0.4876 0.952
('M', 0.1, 'zero (proposed)') -0.0001 0.0002 0.4851 0.9283
('M', 0.3, 'correct') -0.0019 -0.0034 0.4774 0.8943
('M', 0.3, 'mean') -0.0002 -0.0007 0.4835 0.9274
('M', 0.3, 'native') -0.0047 -0.0006 0.4843 0.9514
('M', 0.3, 'zero (proposed)') -0.0004 -0.0006 0.4848 0.9275
('M', 0.5, 'correct') -0.0026 -0.0032 0.4768 0.8945
('M', 0.5, 'mean') -0.0005 -0.0008 0.4832 0.9272
('M', 0.5, 'native') -0.0055 -0.001 0.4835 0.951
('M', 0.5, 'zero (proposed)') -0.0002 -0.0007 0.485 0.9274
('M', 1.0, 'correct') 0.0002 -0.0029 0.4796 0.8948
('M', 1.0, 'mean') -0.0002 -0.0031 0.4835 0.9249
('M', 1.0, 'native') -0 -0.0028 0.4889 0.9493
('M', 1.0, 'zero (proposed)') 0.0001 -0.003 0.4853 0.9251
('N', 0.1, 'GBDT reference') -0.0005 -0.0007 0.4986 0.9821
('N', 0.1, 'correct') -0.0019 -0.0045 0.4774 0.8932
('N', 0.1, 'mean') 0.0001 -0.0008 0.4838 0.9272
('N', 0.1, 'native') -0.0019 -0.0006 0.4871 0.9514
('N', 0.1, 'zero (proposed)') 0.0002 -0.0008 0.4854 0.9274
('N', 0.3, 'GBDT reference') -0.006 -0.002 0.493 0.9808
('N', 0.3, 'correct') -0.0129 -0.0158 0.4665 0.8819
('N', 0.3, 'mean') -0.004 -0.0027 0.4797 0.9253
('N', 0.3, 'native') -0.013 -0.0016 0.476 0.9505
('N', 0.3, 'zero (proposed)') -0.0035 -0.0026 0.4817 0.9255
('N', 0.5, 'GBDT reference') -0.0106 -0.0021 0.4885 0.9807
('N', 0.5, 'correct') -0.0136 -0.0219 0.4658 0.8757
('N', 0.5, 'mean') -0.0034 -0.0039 0.4803 0.9242
('N', 0.5, 'native') -0.0144 -0.0019 0.4745 0.9502
('N', 0.5, 'zero (proposed)') -0.0028 -0.0035 0.4825 0.9246
('N', 1.0, 'GBDT reference') -0.016 -0.004 0.4831 0.9788
('N', 1.0, 'correct') -0.0068 -0.0142 0.4725 0.8834
('N', 1.0, 'mean') -0.005 -0.0063 0.4787 0.9217
('N', 1.0, 'native') -0.0052 -0.0031 0.4837 0.9489
('N', 1.0, 'zero (proposed)') -0.0036 -0.0061 0.4817 0.922
('S', 0.1, 'correct') -0.003 -0.0001 0.4763 0.8976
('S', 0.1, 'mean') -0.0012 0 0.4825 0.928
('S', 0.1, 'native') -0.0009 -0.0009 0.488 0.9512
('S', 0.1, 'zero (proposed)') -0.0012 0 0.484 0.9281
('S', 0.3, 'correct') -0.0045 -0.0002 0.4748 0.8975
('S', 0.3, 'mean') -0.0018 0 0.4819 0.928
('S', 0.3, 'native') -0.0038 -0.0023 0.4852 0.9497
('S', 0.3, 'zero (proposed)') -0.0016 -0 0.4836 0.9281
('S', 0.5, 'correct') -0.0064 -0.0003 0.4729 0.8973
('S', 0.5, 'mean') -0.0025 0 0.4812 0.928
('S', 0.5, 'native') -0.0019 -0.0021 0.4871 0.9499
('S', 0.5, 'zero (proposed)') -0.0026 0 0.4826 0.9282
('S', 1.0, 'correct') -0.0074 -0.0001 0.472 0.8976
('S', 1.0, 'mean') -0.004 -0 0.4798 0.928
('S', 1.0, 'native') -0.0044 -0.0005 0.4845 0.9515
('S', 1.0, 'zero (proposed)') -0.0032 0 0.482 0.9282
('S_sup', 0.1, 'correct') -0.0008 -0.0012 0.4785 0.8965
('S_sup', 0.1, 'mean') 0.0001 -0.0008 0.4839 0.9272
('S_sup', 0.1, 'native') -0.0003 -0.0008 0.4887 0.9512
('S_sup', 0.1, 'zero (proposed)') -0.0001 -0.0008 0.4851 0.9274
('S_sup', 0.3, 'correct') -0.0047 -0.0063 0.4746 0.8914
('S_sup', 0.3, 'mean') -0.0013 -0.0035 0.4825 0.9245
('S_sup', 0.3, 'native') -0.0041 -0.0034 0.4848 0.9486
('S_sup', 0.3, 'zero (proposed)') -0.0025 -0.0036 0.4828 0.9245
('S_sup', 0.5, 'correct') -0.0066 -0.011 0.4727 0.8867
('S_sup', 0.5, 'mean') -0.0014 -0.0057 0.4823 0.9223
('S_sup', 0.5, 'native') -0.0044 -0.0054 0.4846 0.9466
('S_sup', 0.5, 'zero (proposed)') -0.002 -0.0059 0.4832 0.9222
('S_sup', 1.0, 'correct') -0.0013 -0.0125 0.478 0.8852
('S_sup', 1.0, 'mean') -0.0002 -0.0106 0.4835 0.9174
('S_sup', 1.0, 'native') -0.0007 -0.0105 0.4882 0.9415
('S_sup', 1.0, 'zero (proposed)') -0.0008 -0.0109 0.4844 0.9172
('T', 0.1, 'GBDT reference') -0.0028 -0.0008 0.4963 0.982
('T', 0.1, 'correct') -0.002 0.0002 0.4773 0.8979
('T', 0.1, 'mean') -0.0008 0.0001 0.483 0.9282
('T', 0.1, 'native') -0.0126 -0.0002 0.4764 0.9519
('T', 0.1, 'zero (proposed)') -0.0009 0 0.4843 0.9282
('T', 0.3, 'GBDT reference') -0.0057 -0.0015 0.4933 0.9813
('T', 0.3, 'correct') -0.0038 -0.0008 0.4755 0.8969
('T', 0.3, 'mean') -0.0017 -0.0009 0.482 0.9271
('T', 0.3, 'native') -0.0263 -0.002 0.4627 0.95
('T', 0.3, 'zero (proposed)') -0.0015 -0.001 0.4837 0.9272
('T', 0.5, 'GBDT reference') -0.0096 -0.0023 0.4895 0.9805
('T', 0.5, 'correct') -0.0127 -0.0003 0.4667 0.8974
('T', 0.5, 'mean') -0.0063 -0.0006 0.4775 0.9275
('T', 0.5, 'native') -0.0256 -0.0011 0.4634 0.9509
('T', 0.5, 'zero (proposed)') -0.0058 -0.0006 0.4794 0.9276
('T', 1.0, 'GBDT reference') -0.0141 -0.0047 0.485 0.9781
('T', 1.0, 'correct') -0.007 -0.0006 0.4723 0.8971
('T', 1.0, 'mean') -0.0104 -0.0008 0.4734 0.9272
('T', 1.0, 'native') -0.013 -0.002 0.476 0.9501
('T', 1.0, 'zero (proposed)') -0.0104 -0.0007 0.4748 0.9274
('W', 0.1, 'GBDT reference') -0.0068 -0.0625 0.4922 0.9202
('W', 0.1, 'correct') -0.0061 -0.0035 0.4732 0.8942
('W', 0.1, 'mean') -0.0013 -0.0073 0.4825 0.9207
('W', 0.1, 'native') -0.0061 -0.0647 0.4828 0.8873
('W', 0.1, 'zero (proposed)') -0.0014 -0.0079 0.4838 0.9203
('W', 0.3, 'GBDT reference') -0.0107 -0.1242 0.4884 0.8586
('W', 0.3, 'correct') -0.0238 -0.0117 0.4556 0.886
('W', 0.3, 'mean') -0.0042 -0.0213 0.4796 0.9068
('W', 0.3, 'native') -0.0064 -0.1207 0.4826 0.8313
('W', 0.3, 'zero (proposed)') -0.0039 -0.022 0.4814 0.9061
('W', 0.5, 'GBDT reference') -0.0111 -0.1486 0.488 0.8342
('W', 0.5, 'correct') -0.033 -0.0217 0.4463 0.8759
('W', 0.5, 'mean') -0.0088 -0.0351 0.4749 0.8929
('W', 0.5, 'native') -0.0115 -0.1359 0.4775 0.8161
('W', 0.5, 'zero (proposed)') -0.0083 -0.0358 0.4769 0.8924
('W', 1.0, 'GBDT reference') -0.0218 -0.1082 0.4773 0.8746
('W', 1.0, 'correct') -0.0136 -0.0312 0.4658 0.8664
('W', 1.0, 'mean') -0.0148 -0.0603 0.4689 0.8677
('W', 1.0, 'native') -0.0156 -0.0839 0.4733 0.8681
('W', 1.0, 'zero (proposed)') -0.014 -0.0607 0.4712 0.8674
('all', 0.1, 'correct') -0.0163 -0.0147 0.463 0.883
('all', 0.1, 'mean') -0.0018 -0.0114 0.4819 0.9166
('all', 0.1, 'native') -0.0217 -0.0705 0.4673 0.8815
('all', 0.1, 'zero (proposed)') -0.0018 -0.0109 0.4834 0.9173
('all', 0.3, 'correct') -0.0469 -0.0599 0.4324 0.8378
('all', 0.3, 'mean') -0.0149 -0.0389 0.4688 0.8891
('all', 0.3, 'native') -0.0657 -0.1403 0.4232 0.8117
('all', 0.3, 'zero (proposed)') -0.0161 -0.0387 0.4691 0.8895
('all', 0.5, 'correct') -0.0626 -0.107 0.4167 0.7907
('all', 0.5, 'mean') -0.0274 -0.079 0.4564 0.8491
('all', 0.5, 'native') -0.0849 -0.216 0.4041 0.736
('all', 0.5, 'zero (proposed)') -0.0284 -0.0819 0.4568 0.8462