max |direct - evidently| over 80 checks: 0.000000 reference=adult.data 32,561 rows, random windows, R=5000 n mean share p5 p95 dataset rate per-column alarm 50 0.703 0.500 0.875 98.4% Educat=1.00 Marita=0.69 Native=1.00 Occupa=1.00 Race=0.48 Relati=0.63 Sex=0.05 Workcl=0.77 100 0.468 0.250 0.625 56.8% Educat=0.97 Marita=0.24 Native=1.00 Occupa=0.89 Race=0.12 Relati=0.19 Sex=0.01 Workcl=0.32 200 0.234 0.125 0.375 1.3% Educat=0.49 Marita=0.02 Native=1.00 Occupa=0.31 Race=0.01 Relati=0.01 Sex=0.00 Workcl=0.03 500 0.099 0.000 0.125 0.0% Educat=0.00 Marita=0.00 Native=0.79 Occupa=0.00 Race=0.00 Relati=0.00 Sex=0.00 Workcl=0.00 1000 0.000 0.000 0.000 0.0% Educat=0.00 Marita=0.00 Native=0.00 Occupa=0.00 Race=0.00 Relati=0.00 Sex=0.00 Workcl=0.00 2000 0.000 0.000 0.000 0.0% Educat=0.00 Marita=0.00 Native=0.00 Occupa=0.00 Race=0.00 Relati=0.00 Sex=0.00 Workcl=0.00 mean distance per column and the window where the simulated mean falls below 0.1 column k sim50 fml50 sim100 fml100 12.5(k-1) sim<0.1 at Education 16 0.206 0.194 0.146 0.137 188 225 Marital-Status 7 0.119 0.122 0.084 0.087 75 70 Native_country 42 0.201 0.320 0.176 0.226 512 600 Occupation 15 0.192 0.187 0.132 0.132 175 175 Race 5 0.104 0.100 0.075 0.071 50 55 Relationship 6 0.115 0.112 0.077 0.079 62 65 Sex 2 0.040 0.050 0.028 0.035 12 10 Workclass 9 0.133 0.141 0.090 0.100 100 85 reference=missing rows dropped 30,162 rows: distribution of the share over random windows, R=5000 n=50 mean=0.691 dataset rate=97.5% 0.250:0.1% 0.375:2.4% 0.500:13.4% 0.625:31.4% 0.750:35.8% 0.875:15.5% 1.000:1.4% n=100 mean=0.452 dataset rate=51.1% 0.125:0.2% 0.250:8.4% 0.375:40.3% 0.500:35.4% 0.625:12.9% 0.750:2.5% 0.875:0.3% column alarm rate: simulated vs chi-square prediction P(chi2(k-1) >= 8 n t^2), t=0.1 column k sim50 fml50 sim100 fml100 sim200 fml200 sim 5% at fml 5% at Education 16 100% 100% 97% 92% 50% 38% 350 312 Marital-Status 7 70% 68% 24% 24% 2% 1% 170 157 Native_country 42 100% 100% 100% 100% 100% 100% 825 712 Occupation 15 100% 100% 89% 89% 32% 31% 300 296 Race 5 48% 41% 13% 9% 1% 0% 140 119 Relationship 6 61% 55% 20% 16% 1% 1% 150 138 Sex 2 5% 5% 1% 0% 0% 0% 50 48 Workclass 9 77% 86% 32% 43% 3% 4% 180 194 window 100 over one random order of the rows: sliding by one row, then non-overlapping windows sliding: alarm 55.4%, next window also alarms 94.6% 3 in a row: sliding 49.9%, non-overlapping 16.4% 5 in a row: sliding 45.9%, non-overlapping 1.9% 10 in a row: sliding 38.2%, non-overlapping 0.0% size of a distance of 0.1 50:50 vs 64:36 -> 0.1 a 2% category disappears -> 0.084 a 5% category disappears -> 0.133 no overlap -> 0.833