---
title: "Infogram Train Subset Models Demo Notebook"
output:
  html_document:
    df_print: paged
---

```{r}
library(h2o)

h2o.init()

# Import HMDA dataset
f <- "https://erin-data.s3.amazonaws.com/admissible/data/hmda_lar_2018_sample.csv"
col_types <- list(by.col.name = c("high_priced"),
                  types = c("factor"))
df <- h2o.importFile(path = f, col.types = col_types)

splits <- h2o.splitFrame(df, ratios = 0.8, seed = 1)
train <- splits[[1]]
test <- splits[[2]]

# Response column and predictor columns
y <- "high_priced"
x <- c("loan_amount",
       "loan_to_value_ratio",
       "loan_term",
       "intro_rate_period",
       "property_value",
       "income",
       "debt_to_income_ratio")

# Fairness related information
protected_columns <- c("derived_race", "derived_sex")
reference <- c("White", "Male")
favorable_class <- "0"

# Infogram
ig <- h2o.infogram(y = y, x = x, training_frame = train, protected_columns = protected_columns)
plot(ig)

# Admissible score frame
asf <- ig@admissible_score
asf
```

```{r}
da <- h2o.no_progress(h2o.infogram_train_subset_models(ig, h2o.automl, train, test, y, protected_columns, reference, favorable_class, max_models = 10, seed = 1))
da
```

```{r}
pf <- h2o.pareto_front(da, x_metric = "auc", y_metric = "significant_air_min", optimum = "top right", color = "algo")
plot(pf)
pf@pareto_front
```

```{r}
potentially_fair_model <- h2o.getModel(da[da$significant_air_min > 0.8, "model_id"][[1]])
```

```{r fig.height=6, fig.width=12, warning=FALSE, results="asis"}
h2o.inspect_model_fairness(potentially_fair_model, test, protected_columns, reference, favorable_class)
```