cfpb complaints - reticulate

cfpb-complaints.Rmd

---
name: "Ryan Wesslen"
title: "cfpb complaints/reticulate"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
library(reticulate); library(tidyverse)
```

## R Markdown

Let's use pandas to load the original data.

```{python data}
import pandas as pd

data = pd.read_csv('complaints.csv', engine='python')
data.head(n=5)
```

Let's keep only three of the products.

```{python product-clean}
products = ['Credit card','Checking or savings account','Mortgage']

df = data.loc[data["Product"].isin(products)]
```

Let's now remove rare sub-products 

```{python sub-product}
subproducts = ['CD (Certificate of Deposit)','Personal line of credit','Reverse mortgage','VA mortgage']

df = df.loc[~data["Sub-product"].isin(subproducts)]

h = df.groupby(['Sub-product'])['Sub-product'].count()

print(h)
```

```{python sub-product2}
df.loc[df['Sub-product'] == "Home equity loan or line of credit (HELOC)", 'Sub-product'] = "Home equity loan or line of credit"
df.loc[df['Sub-product'] == "Other type of mortgage", 'Sub-product'] = "Other mortgage"

# give 
df.loc[df['Sub-product'] == "None", 'Sub-product'] = "Other banking product or service"
df.loc[df['Product'] == "Credit card", 'Sub-product'] = "Credit card"
```

# Product-Issue Hiearchcy

```{r}
library(collapsibleTree)

library(RColorBrewer)

py$df %>%
  count(Product, `Sub-product`, Issue, `Sub-issue`) %>%
  collapsibleTree(
    c("Product", "Sub-product", "Issue", "Sub-issue"),
    root = "Total",
    tooltip = TRUE,
    attribute = "n",
    nodeSize = 'leafCount',
    width = 800, 
    zoomable = FALSE
  )
```


```{python}
h = df.groupby(['Issue'])['Issue'].count()

print(h)
```


```{r}
py$df %>%
  write_csv(t)
```


```{r eval=FALSE}
py$df %>%
  count(Product) %>%
  arrange(desc(n)) %>%
  ggplot(aes(x = forcats::fct_reorder(Product, count), y = count)) +
  geom_col() +
  coord_flip() +
  ggthemes::theme_economist() +
  labs(x = " ", y = "Number of Complaints", title = "CFPB Complaints by product")
```


## Product Only

```{python partition, echo=FALSE, eval=FALSE}
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(df['Consumer complaint narrative'], df['Product'], random_state = 0)
```

```{python fit-tsne}
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.pipeline import Pipeline
from sklearn.manifold import TSNE
from sklearn.decomposition import TruncatedSVD

text_tsne = Pipeline([
    ('tfidf', TfidfVectorizer()),
    ('svd', TruncatedSVD(n_components=50)),
    ('tsne', TSNE(n_components=2)),
])

# change these!
points = text_tsne.fit_transform(df['Consumer complaint narrative'])
targets = df['Product']
tsne_df = pd.DataFrame({'x':points[:,0],'y':points[:,1],'label':targets})
```


## Run R on it

```{r}
library(tidyverse); library(rsample)

split <- py$tsne_df %>%
  mutate(label = as.factor(label)) %>%
  initial_split()
train_data <- training(split)
test_data <- testing(split)

train_data %>%
  ggplot(aes(x = x, y = y)) +
  geom_point(aes(fill = label), color = "black", size = 0.7, shape = 21) +
  theme_bw() +
  theme(legend.position = "bottom")
```








```{r}
# Learn model
rf = randomForest::randomForest(label ~ x + y, data = train_data, ntree=100)
train_data$predicted = predict(rf, train_data)
test_data$predicted = predict(rf, test_data)
```

```{r}
# Define range of set
lower_x1 = -90
upper_x1 = 90
lower_x2 = -90
upper_x2 = 90
n_grid = 100
# The decision boundaries
grid_x1 = seq(from=lower_x1, to=upper_x1, length.out=n_grid)
grid_x2 = seq(from=lower_x2, to=upper_x2, length.out=n_grid)
grid_df = expand.grid(x = grid_x1, y = grid_x2)
grid_df$predicted = as.character(predict(rf, newdata = grid_df))

prob <- as.data.frame(predict(rf, newdata = grid_df, type = "prob"))
grid_df$prob <- apply(prob, 1, FUN=max)
```

```{r}
library(yardstick)

accuracy(test_data, label, predicted)
```

```{r}
conf_mat(test_data, label, predicted)
```

```{r}
bal_accuracy(test_data, label, predicted)
```

```{r}
ggplot(grid_df, aes(x = x, y = y, fill = predicted)) +
  #geom_hex(alpha=0.4) +
  geom_raster(aes(alpha = prob), interpolate=TRUE) + #alpha = 0.3 to be fixed pred
  geom_point(data = test_data, aes(fill = label), color = "black", shape=21, size = 0.5) +
  theme(legend.position = "right") +
  labs(x = "x1", y = "x2", title = "Random Forest on t-SNE reduced")
```