EDA minimale - Criteo Uplift (dev) Cette EDA vérifie les distributions, les taux principaux et l'équilibre des
covariables avant modélisation. Le traitement principal reste
treatment. exposure est affichée uniquement comme variable
post-traitement diagnostique.
variante : official Criteo v2.1 CSV streaming_sample=1000000 ;
loader : https://go.criteo.net/criteo-research-uplift-v2.1.csv.gz ;
lignes chargées : 1000000 ;
features : f0, f1, f2, f3, f4, f5, f6, f7, f8, f9, f10, f11.
column
rate
n_positive
n
treatment
0.850411
850411
1000000
exposure
0.030457
30457
1000000
visit
0.046868
46868
1000000
conversion
0.002882
2882
1000000
Outcomes par groupe traité / contrôle
outcome
group
treatment_value
n
outcome_rate
visit
treated
1
850411
0.04858709494585559
visit
control
0
149589
0.03709497356089017
conversion
treated
1
850411
0.0030679283311245975
conversion
control
0
149589
0.0018250005013737641
Plus forts déséquilibres de covariables
feature
treated_mean
control_mean
standardized_mean_difference
f3
4.169382824914724
4.237058588582021
-0.05220373520489931
f6
-4.183370219548571
-3.997169211835693
-0.04119850728463768
f5
4.027362865049725
4.0401872633143
-0.030987962453181146
f9
16.061797739096434
15.860110515519786
0.02918450320720227
f1
10.070367585908626
10.067510567448913
0.028848746301362867
f8
3.9332991534647177
3.934715571460335
-0.025239917981684745
f7
5.102543183581748
5.079949151834541
0.019058007614345205
f10
5.333894649174358
5.3317468357104865
0.01288379698077444
Distribution des features
feature
mean
std
min
5%
50%
95%
max
missing_rate
f0
19.62522496565912
5.377089472165765
12.616364906986496
12.616364906986496
21.925550775069965
26.31227777293766
26.745253093826705
0.0
f1
10.069940207374248
0.10484685963397183
10.059654474774549
10.059654474774549
10.059654474774549
10.059654474774549
15.640628400304251
0.0
f2
8.446707061687537
0.2994413048242057
8.214382844395335
8.214382844395335
8.214382844395335
9.004475753846545
9.05195860467341
0.0
f3
4.1795063747259515
1.3358887937595036
-6.134914697830531
0.9738407885786832
4.679881620097284
4.679881620097284
4.679881620097284
0.0
f4
10.339258213183035
0.3447562940961129
10.280525225748212
10.280525225748212
10.280525225748212
10.280525225748212
21.123507734315055
0.0
f5
4.029281253961727
0.42913572725876525
-8.91220852639907
3.0130643418630596
4.115453421277861
4.115453421277861
4.115453421277861
0.0
f6
-4.155516597005807
4.5790470043351
-29.46764030128149
-13.353454565674014
-2.411114577488788
0.294442711255606
0.294442711255606
0.0
f7
5.099163364966716
1.199989143164773
4.833814577796811
4.833814577796811
4.833814577796811
5.868857280286607
11.998396508915
0.0
f8
3.933511034016265
0.05670189159843879
3.63510661177053
3.805965965081211
3.971857989350875
3.971857989350875
3.971857989350875
0.0
f9
16.031627549008828
7.027175641732452
13.190055934673358
13.190055934673358
13.190055934673358
33.71255596226026
75.29501735734412
0.0
f10
5.333573359906111
0.168536956430928
5.300374864042156
5.300374864042156
5.300374864042156
5.300374864042156
6.473915055850751
0.0
f11
-0.17098105418951298
0.023010020719763442
-1.2177237223033317
-0.1686792210005612
-0.1686792210005612
-0.1686792210005612
-0.1686792210005612
0.0
Les features modèle sont limitées à f0-f11.
treatment, exposure, visit et conversion ne sont pas utilisées comme
features.Le groupe traité est majoritaire ; les métriques uplift doivent donc utiliser
des estimations adaptées traité/contrôle, pas l'accuracy.
Les features sont anonymisées : l'analyse segmentaire métier devra rester
prudente.
reports/tables/eda_binary_rates_dev_visit.csvreports/tables/eda_outcome_rates_by_treatment_dev_visit.csvreports/tables/eda_feature_balance_by_treatment_dev_visit.csvreports/tables/eda_feature_quantile_summary_dev_visit.csvreports/figures/outcome_rates_by_treatment_dev_visit.pngreports/figures/feature_distributions_sample_dev_visit.pngreports/figures/feature_balance_by_treatment_dev_visit.png