{ "claim2": { "normalisation": "percent of the largest ERT over all methods AND all numbers of test samples, per dataset and repetition, then averaged (Table 2 caption); negative ERT clipped to 0", "paper_table2_L1": { "CheapBetterLGBMClassifier": 68.4, "PartitionWise": 38.3, "BetterCatBoost": 68.7, "RF": 65.9, "XT": 65.9, "tabICL": 71.9, "TabPFN": 71.6 }, "variants": { "v1 (results_old.csv)": { "rows": 2800, "L1_percent_of_max": { "CheapBetterLGBMClassifier": 68.03638140127894, "XT": 65.52544651991043, "RF": 65.39950342632956, "tabICL": 71.5102160122718, "PartitionWise": 37.8207068825266, "TabPFN": 71.23955924079642, "BetterCatBoost": 68.66960392756035 }, "lightgbm": 68.03638140127894, "partitionwise": 37.8207068825266, "lightgbm_abs_err_vs_paper": 0.36361859872106095, "partitionwise_abs_err_vs_paper": 0.47929311747339653, "ordering_lightgbm_above_partitionwise": true }, "v2 (results.csv)": { "rows": 5600, "L1_percent_of_max": { "BetterCatBoost": 70.31627805977823, "CheapBetterLGBMClassifier": 68.92900479082506, "XT": 66.53143321909444, "RF": 67.2767890259383, "tabICL": 72.66607482837067, "PartitionWise": 33.670548352764385, "TabPFN": 72.13568950693384 }, "lightgbm": 68.92900479082506, "partitionwise": 33.670548352764385, "lightgbm_abs_err_vs_paper": 0.5290047908250557, "partitionwise_abs_err_vs_paper": 4.629451647235612, "ordering_lightgbm_above_partitionwise": true } }, "other_metrics_v2": { "L2": { "BetterCatBoost": 49.0071922409777, "CheapBetterLGBMClassifier": 46.435713430066265, "XT": 30.302556102139143, "RF": 36.87651319709627, "tabICL": 53.9583162881316, "PartitionWise": 10.530135474406409, "TabPFN": 49.748553099895055 }, "KL": { "BetterCatBoost": 54.05819133012458, "CheapBetterLGBMClassifier": 50.935353447465104, "XT": 27.670695250125288, "RF": 35.0531315347866, "tabICL": 59.246746583269534, "PartitionWise": 11.012516952290087, "TabPFN": 55.22562054445726 } } }, "claim5": { "n_rows": 80, "decomposition_max_residual": 8.940696721859709e-08, "per_dataset_method": { "MNIST|likelihood": { "KL": [ -0.18710587758504865, 0.006658818037739759 ], "KL_plus": [ -0.12831058113152882, 0.005195663786398997 ], "KL_minus": [ -0.05879531943273254, 0.0024970636507724605 ] }, "MNIST|cumulative": { "KL": [ -0.21624940017182154, 0.01666774016886198 ], "KL_plus": [ -0.15939179494622038, 0.01161647543368673 ], "KL_minus": [ -0.05685761234871077, 0.005610714577396073 ] }, "FashionMNIST|likelihood": { "KL": [ -0.06811388991773122, 0.007993181559542269 ], "KL_plus": [ -0.04212642349302763, 0.005677110447665139 ], "KL_minus": [ -0.02598746716976159, 0.004767266311988138 ] }, "FashionMNIST|cumulative": { "KL": [ -0.2601063385605812, 0.013834332625820572 ], "KL_plus": [ -0.18529993444681164, 0.010265827845468862 ], "KL_minus": [ -0.07480641230940815, 0.003877947026407775 ] }, "CIFAR10|likelihood": { "KL": [ 0.02832810527194678, 0.00618225335103705 ], "KL_plus": [ 0.006766273948223831, 0.001416814412288527 ], "KL_minus": [ 0.02156181978604231, 0.007053554145809787 ] }, "CIFAR10|cumulative": { "KL": [ -0.017296617239686982, 0.007835687912677925 ], "KL_plus": [ -0.02951636482503982, 0.005342355728146586 ], "KL_minus": [ 0.01221973485697023, 0.0061652773637478776 ] }, "CIFAR100|likelihood": { "KL": [ 0.4086424125659942, 0.02541961045936785 ], "KL_plus": [ 0.08522980659590909, 0.021734443672694895 ], "KL_minus": [ 0.3234126213317108, 0.01978366305585833 ] }, "CIFAR100|cumulative": { "KL": [ 0.1905506563616333, 0.024133401088240462 ], "KL_plus": [ 0.016023607273191902, 0.007820076340537635 ], "KL_minus": [ 0.1745270594703445, 0.026338485382953438 ] } }, "n_cells": 8, "n_cells_with_divergent_components": 8, "naming": "the released columns are 'underconfident'/'overconfident'; an underconfident predictor makes sets that are too wide, i.e. over-coverage, which is the paper's l_+ component" }, "claim1": { "alpha": 0.1, "d": 8, "n_test": 5000, "n_cal": 3000, "seeds": 5, "standard": { "L1": 0.063272, "true_L1": 0.0702964914948382, "L2": 0.005256268125075593, "true_L2": 0.007167963207236776, "KL": 0.030604667681682535, "true_KL": 0.0416029844930876, "marginal_coverage": 0.89856 }, "oracle": { "L1": 0.0011679999999999982, "true_L1": 0.0, "L2": -0.00017319047147024304, "true_L2": 0.0, "KL": -0.0010198970862035728, "true_KL": 0.0, "marginal_coverage": 0.90032 }, "oracle_L1_near_zero": true, "standard_L1_positive": true, "separation_L1": 0.062104 }, "claim6": { "true": 0.0690035899522276, "in_sample_tree": 0.1743, "cv_tree": 0.021024999999999995, "cv_hgb": 0.05347500000000001, "oracle_in_sample_tree": 0.17670000000000002, "oracle_cv_tree": 0.004424999999999995, "n_test": 1000, "in_sample_inflation_factor": 8.290130796670633, "oracle_in_sample_false_positive": 0.17670000000000002, "oracle_cv_is_near_zero": true }, "claim4": { "n_test": 6000, "seeds": 8, "scenarios": { "conservative (radius x1.35)": { "L1_plus_est": 0.06408333333333331, "L1_minus_est": 0.0010625000000000014, "L1_plus_true": 0.06608391764460793, "L1_minus_true": 0.003515496941181244, "KL_plus_est": 0.044628850516583406, "KL_minus_est": 0.00021369709520476876, "KL_plus_true": 0.051062027112734255, "KL_minus_true": 0.0011484626555137195, "marginal_coverage": 0.9630208333333333, "scale": 1.35, "L1_ratio_plus_over_minus": 60.31372549019598, "dominant_side_est": "plus", "dominant_side_true": "plus", "side_attributed_correctly": true }, "aggressive (radius x0.75)": { "L1_plus_est": 0.008629166666666663, "L1_minus_est": 0.11244166666666666, "L1_plus_true": 0.012156084075637029, "L1_minus_true": 0.11614065781185745, "KL_plus_est": 0.0021941574834898414, "KL_minus_est": 0.08249845596638003, "KL_plus_true": 0.006623309264618229, "KL_minus_true": 0.09067184559363284, "marginal_coverage": 0.7961875, "scale": 0.75, "L1_ratio_plus_over_minus": 0.0767434966278811, "dominant_side_est": "minus", "dominant_side_true": "minus", "side_attributed_correctly": true }, "standard (radius x1)": { "L1_plus_est": 0.029933333333333322, "L1_minus_est": 0.03262083333333334, "L1_plus_true": 0.03334231184161376, "L1_minus_true": 0.03734773786332873, "KL_plus_est": 0.01555761941505708, "KL_minus_est": 0.013312671224877568, "KL_plus_true": 0.02204298832902235, "KL_minus_true": 0.01938098116174198, "marginal_coverage": 0.8973125000000001, "scale": 1.0, "L1_ratio_plus_over_minus": 0.917613999233618, "dominant_side_est": "minus", "dominant_side_true": "minus", "side_attributed_correctly": true } }, "all_sides_attributed_correctly": true, "separation_conservative_vs_aggressive": 785.913180144099 }, "claim3": { "rows": [ { "ert_std": 0.03576, "ert_ora": -0.0016800000000000014, "cg_std": 0.031072781526658745, "cg_ora": 0.031199778923785264, "true_std": 0.06880447557795312, "n_test": 500, "ert_error_vs_true": 0.03304447557795312, "covgap_error_vs_true": 0.037731694051294375, "ert_separation": 0.03744, "covgap_separation": -0.0001269973971265194 }, { "ert_std": 0.05286, "ert_ora": -0.0035400000000000015, "cg_std": 0.025883724155466782, "cg_ora": 0.024059037379189192, "true_std": 0.06889486259000807, "n_test": 1000, "ert_error_vs_true": 0.016034862590008074, "covgap_error_vs_true": 0.04301113843454129, "ert_separation": 0.0564, "covgap_separation": 0.0018246867762775898 }, { "ert_std": 0.05676000000000001, "ert_ora": -0.002210000000000001, "cg_std": 0.018405934181208283, "cg_ora": 0.014972773544179094, "true_std": 0.06954455557994782, "n_test": 2000, "ert_error_vs_true": 0.012784555579947808, "covgap_error_vs_true": 0.05113862139873954, "ert_separation": 0.058970000000000015, "covgap_separation": 0.003433160637029189 }, { "ert_std": 0.06160800000000001, "ert_ora": -0.0005080000000000013, "cg_std": 0.014439326900402834, "cg_ora": 0.010073670837015896, "true_std": 0.06944134707377626, "n_test": 5000, "ert_error_vs_true": 0.007833347073776253, "covgap_error_vs_true": 0.055002020173373425, "ert_separation": 0.06211600000000001, "covgap_separation": 0.004365656063386939 } ], "n_seeds": 10, "ert_closer_to_truth_at_all_n": true, "separation_ratio_at_max_n": 14.228331114066169 } }