import h2o
from h2o.automl import *
from h2o.estimators.infogram import H2OInfogram

h2o.connect()

# Import HDMA dataset
f = "https://erin-data.s3.amazonaws.com/admissible/data/hmda_lar_2018_sample.csv"
col_types = {'high_priced': "enum"}
df = h2o.import_file(path=f, col_types=col_types)

# We will split the data so that we can test/compare performance
# of admissible vs non-admissible models later
train, test = df.split_frame(ratios=[0.8], seed=1)

# Response column and predictor columns
y = "high_priced"
x =  ["loan_amount",
      "loan_to_value_ratio",
      "loan_term",
      "intro_rate_period",
      "property_value",
      "income",
      "debt_to_income_ratio"]

# Fairness related information
protected_columns = ["derived_race", "derived_sex"]
reference = ["White", "Male"]
favorable_class = "0"

# Infogram
ig = H2OInfogram(protected_columns=protected_columns)
ig.train(y=y, x=x, training_frame=train)
ig.plot()

# Admissible score frame
asf = ig.get_admissible_score_frame()
asf

Connecting to H2O server at http://localhost:54321 ... successful.
Warning: Version mismatch. H2O is version 3.39.0.99999, but the h2o-python package is version 0.0.local. This is a developer build, please contact your developer.

H2O_cluster_uptime:	1 hour 43 mins
H2O_cluster_timezone:	Europe/Prague
H2O_data_parsing_timezone:	UTC
H2O_cluster_version:	3.39.0.99999
H2O_cluster_version_age:	1 hour and 47 minutes
H2O_cluster_name:	tomasfryda
H2O_cluster_total_nodes:	1
H2O_cluster_free_memory:	892 Mb
H2O_cluster_total_cores:	16
H2O_cluster_allowed_cores:	8
H2O_cluster_status:	locked, healthy
H2O_connection_url:	http://localhost:54321
H2O_connection_proxy:	{"http": null, "https": null}
H2O_internal_security:	False
Python_version:	3.9.15 final

Parse progress: |████████████████████████████████████████████████████████████████| (done) 100%
infogram Model Build progress: |█████████████████████████████████████████████████| (done) 100%

column	admissible	admissible_index	relevance_index	safety_index	cmi_raw
loan_to_value_ratio	1	1	1	1	0.0941406
property_value	1	0.232405	0.145431	0.294744	0.0277474
loan_amount	1	0.176956	0.122526	0.218206	0.0205421
income	0	0.106802	0.0265004	0.148697	0.0139985
intro_rate_period	0	0.0627864	0.0551325	0.0696038	0.00655254
loan_term	0	0.046236	0.0550699	0.0352539	0.00331882
debt_to_income_ratio	0	0.0402932	0.0537012	0.0190596	0.00179428

[7 rows x 6 columns]

# Train models on different subsets of features selected by infogram
da = ig.train_subset_models(H2OAutoML, y, train, test, protected_columns, reference, favorable_class, max_models=10, seed=1)
da

AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
AutoML progress: |███████████████████████████████████████████████████████████████| (done) 100%
Parse progress: |████████████████████████████████████████████████████████████████| (done) 100%

model_id	auc	logloss	aucpr	mean_per_class_error	rmse	mse	training_time_ms	predict_time_per_row_ms	algo	num_of_features	var	corrected_var	air_min	air_mean	air_median	air_max	cair	significant_air_min	significant_air_mean	significant_air_median	significant_air_max	p.value_min	p.value_mean	p.value_median	p.value_max
StackedEnsemble_AllModels_1_AutoML_35_20221208_180609	0.85075	0.271505	0.47292	0.252432	0.285302	0.0813975	22582	0.029297	StackedEnsemble	7	0.00529065	0.00529065	0.765944	0.97949	1.00099	1.11694	0.990789	0.765944	0.925672	0.860042	1.11694	1.10185e-39	0.432293	0.455291	1
StackedEnsemble_BestOfFamily_1_AutoML_35_20221208_180609	0.849025	0.272757	0.46821	0.259026	0.285825	0.081696	24597	0.018246	StackedEnsemble	7	0.00635303	0.00635303	0.733577	0.97663	1.00147	1.13548	0.990452	0.733577	0.914688	0.850797	1.13548	1.62109e-45	0.478835	0.497316	1
StackedEnsemble_AllModels_1_AutoML_34_20221208_180217	0.848842	0.272335	0.474325	0.250286	0.285504	0.0815126	22722	0.022641	StackedEnsemble	6	0.00594179	0.00594179	0.745615	0.979764	1.00394	1.12859	0.988746	0.745615	0.907798	0.844812	1.12859	2.36648e-44	0.494873	0.57807	1
StackedEnsemble_BestOfFamily_1_AutoML_34_20221208_180217	0.848441	0.272818	0.472626	0.251506	0.285672	0.0816086	21627	0.018372	StackedEnsemble	6	0.00511355	0.00511355	0.737653	0.967604	1.00079	1.12827	0.988942	0.737653	0.904503	0.842256	1.12827	1.70464e-42	0.463614	0.586117	1
XGBoost_3_AutoML_35_20221208_180609	0.847691	0.273393	0.46894	0.263194	0.286006	0.0817993	3908	0.002935	XGBoost	7	0.00593699	0.00593699	0.761509	0.977403	1.00216	1.11262	0.990231	0.761509	0.921565	0.8482	1.11262	1.65569e-42	0.391109	0.532732	1
XGBoost_3_AutoML_34_20221208_180217	0.846846	0.27358	0.47131	0.247711	0.28601	0.0818018	3355	0.002503	XGBoost	6	0.0062938	0.0062938	0.713172	0.972281	1.00066	1.15728	0.988252	0.713172	0.897472	0.845538	1.11881	9.8231e-44	0.437499	0.392908	1
XGBoost_2_AutoML_35_20221208_180609	0.845576	0.275135	0.461429	0.249549	0.28729	0.0825357	3830	0.00315	XGBoost	7	0.00535889	0.00535889	0.775897	0.980804	1.00512	1.10141	0.989421	0.775897	0.937747	0.955149	1.10141	2.16778e-40	0.421098	0.560022	1
StackedEnsemble_AllModels_1_AutoML_33_20221208_175909	0.844664	0.275223	0.463859	0.255883	0.28689	0.0823061	19782	0.027434	StackedEnsemble	5	0.00572474	0.00572474	0.736285	0.973376	1.00143	1.12269	0.988634	0.736285	0.910441	0.842579	1.12269	1.10785e-46	0.432894	0.455854	1
StackedEnsemble_BestOfFamily_1_AutoML_33_20221208_175909	0.843951	0.275925	0.458906	0.256599	0.287187	0.0824762	18709	0.016466	StackedEnsemble	5	0.00548182	0.00548182	0.759505	0.976403	1.0103	1.11621	0.989056	0.759505	0.910667	0.850645	1.11621	2.56641e-42	0.4037	0.310036	1
GBM_4_AutoML_34_20221208_180217	0.843922	0.276234	0.45389	0.25621	0.28764	0.0827368	2254	0.007186	GBM	6	0.00581139	0.00581139	0.796127	0.979782	1.00007	1.11196	0.992499	0.796127	0.953822	0.961102	1.11196	3.97915e-33	0.481223	0.685476	1

[84 rows x 26 columns]

# Select the best models with respect to AUC and AIR
h2o.explanation.pareto_front(da, x_metric="auc", y_metric="significant_air_min", optimum="top right", color_col="algo")

model_id	auc	logloss	aucpr	mean_per_class_error	rmse	mse	training_time_ms	predict_time_per_row_ms	algo	num_of_features	var	corrected_var	air_min	air_mean	air_median	air_max	cair	significant_air_min	significant_air_mean	significant_air_median	significant_air_max	p.value_min	p.value_mean	p.value_median	p.value_max
StackedEnsemble_AllModels_1_AutoML_35_20221208_180609	0.85075	0.271505	0.47292	0.252432	0.285302	0.0813975	22582	0.029297	StackedEnsemble	7	0.00529065	0.00529065	0.765944	0.97949	1.00099	1.11694	0.990789	0.765944	0.925672	0.860042	1.11694	1.10185e-39	0.432293	0.455291	1
XGBoost_2_AutoML_35_20221208_180609	0.845576	0.275135	0.461429	0.249549	0.28729	0.0825357	3830	0.00315	XGBoost	7	0.00535889	0.00535889	0.775897	0.980804	1.00512	1.10141	0.989421	0.775897	0.937747	0.955149	1.10141	2.16778e-40	0.421098	0.560022	1
GBM_4_AutoML_34_20221208_180217	0.843922	0.276234	0.45389	0.25621	0.28764	0.0827368	2254	0.007186	GBM	6	0.00581139	0.00581139	0.796127	0.979782	1.00007	1.11196	0.992499	0.796127	0.953822	0.961102	1.11196	3.97915e-33	0.481223	0.685476	1
XGBoost_1_AutoML_35_20221208_180609	0.843448	0.277989	0.454395	0.254712	0.289126	0.0835941	5229	0.003437	XGBoost	7	0.00351762	0.00351762	0.799487	0.984814	1.00626	1.09772	0.990846	0.799487	0.929281	0.880262	1.09772	2.0444e-34	0.455747	0.516162	1
XGBoost_1_AutoML_33_20221208_175909	0.837357	0.28159	0.44418	0.254342	0.29065	0.0844775	4686	0.002913	XGBoost	5	0.00485077	0.00485077	0.807207	0.975157	1.00267	1.12213	0.99275	0.807207	0.937024	0.878041	1.09295	5.95856e-31	0.331811	0.277824	1

[5 rows x 26 columns]

# Pick a model that satisfies the four fifths rule
potentially_fair_model = h2o.get_model(da[da["significant_air_min"] > 0.8, "model_id"][0, 0])

# Inspect the model
potentially_fair_model.inspect_model_fairness(test, protected_columns, reference, favorable_class);

Overview for model XGBoost_1_AutoML_33_20221208_175909

The following table shows fairness metrics for intersections determined using the protected_columns. Apart from the fairness metrics, there is a p-value from Fisher's exact test or G-test (depends on the size of the intersections) for hypothesis that being selected (positive response) is independent to being in the reference group or a particular protected group. After the table there are two kinds of plot. The first kind starts with AIR prefix which stands for Adverse Impact Ratio. These plots show values relative to the reference group and also show two dashed lines corresponding to 0.8 and 1.25 (the four-fifths rule). The second kind is showing the absolute value of given metrics. The reference group is shown by using a different colored bar.

derived_race	derived_sex	AIR_auc	AIR_aucpr	AIR_f1	AIR_selectedRatio	auc	aucpr	f1	p.value	selectedRatio	total
2 or more minority races	Female	1.02182	1.00685	0.967116	0.95515	0.857143	0.979769	0.888889	0.72182	0.8125	16
American Indian or Alaska Native	Female	0.90449	0.974675	0.897605	0.852864	0.758721	0.948459	0.825	0.0179482	0.72549	51
Asian	Female	1.09238	1.01225	1.04543	1.05406	0.916327	0.985018	0.960864	0.00869533	0.896635	416
Black or African American	Female	0.896947	0.921109	0.888481	0.807207	0.752393	0.896334	0.816614	5.95856e-31	0.686652	884
Native Hawaiian or Other Pacific Islander	Female	1.10847	1.01681	1.03361	1.12213	0.929825	0.989465	0.95	0.236259	0.954545	22
Race Not Available	Female	0.939426	0.99119	1.00149	1.02115	0.788027	0.964529	0.920482	0.318718	0.868644	472
White	Female	0.994853	0.998371	1.00071	1.00535	0.834521	0.971517	0.91977	0.477256	0.855198	5290
2 or more minority races	Male	0.638638	0.921521	0.92062	0.881677	0.535714	0.896734	0.846154	0.283053	0.75	16
American Indian or Alaska Native	Male	0.972198	0.980643	0.982094	0.950828	0.815517	0.954266	0.902655	0.309009	0.808824	68
Asian	Male	1.13647	1.02467	1.05459	1.09295	0.953313	0.997109	0.969287	8.73843e-12	0.929724	868
Black or African American	Male	0.894786	0.944882	0.916299	0.878041	0.750581	0.919467	0.842182	2.56071e-14	0.746907	889
Native Hawaiian or Other Pacific Islander	Male	0.88448	0.956978	1.01781	1.0123	0.741935	0.931238	0.935484	1	0.861111	36
Race Not Available	Male	1.02716	1.01154	1.00882	1.0185	0.861626	0.984328	0.927215	0.272596	0.866385	711
White	Male	1	1	1	1	0.838838	0.973102	0.919113	1	0.850652	8825

[14 rows x 12 columns]

ROC

The following plot shows a Receiver Operating Characteristic (ROC) for each intersection. This plot could be used for selecting different threshold of the classifier to make it more fair in some sense this is described in, e.g., HARDT, Moritz, PRICE, Eric and SREBRO, Nathan, 2016. Equality of Opportunity in Supervised Learning. arXiv:1610.02413.

Precision-Recall Curve

The following plot shows a Precision-Recall Curve for each intersection.

Permutation Variable Importance

Permutation variable importance is obtained by measuring the distance between prediction errors before and after a feature is permuted; only one feature at a time is permuted.

Variable Importances

Variable	Relative Importance	Scaled Importance	Percentage
loan_to_value_ratio	0.2429555	1.0	0.6288179
property_value	0.0581869	0.2394962	0.1505995
loan_amount	0.0444545	0.1829738	0.1150572
intro_rate_period	0.0258223	0.1062841	0.0668333
income	0.0149494	0.0615314	0.0386920

Partial Dependence Plots for Individual Protected Groups

The following plots show partial dependence for each intersection separately. This plot can be used to see how the membership to a particular intersection influences the dependence on a given feature.

/Users/tomasfryda/sources/h2o-3/h2o-py/h2o/model/extensions/fairness.py:133: RuntimeWarning: All-NaN axis encountered
  maxes.append(np.nanmax(pd[column.lower()]))

SHAP for Individual Protected Groups

The following plots show SHAP contributions for individual intersections and one feature at a time.This plot can be used to see how the membership to a particular intersection influences the dependence on a given feature.

/Users/tomasfryda/sources/h2o-3/h2o-py/h2o/model/extensions/fairness.py:356: RuntimeWarning: All-NaN slice encountered
  maxes.append(np.nanmax(vals))