model_building-v5.Rmd

---
title: "Model Building v5"
output: html_notebook
---

```{r setup, include=F}
rm(list = ls())
load("data/data_preparation.RData")
```

```{r, message=FALSE}
library(keras)
library(dplyr)
library(magrittr)
library(caret)
library(pROC)
```

## 1. Spliting Dataset into Training and Test

```{r echo = F}
set.seed(2019)

train_index <- createDataPartition(
  factor(invoiced),
  p = 0.8,
  list = F,
  times = 1
)
```

### 1.1 Splitting `Invoiced` into Training, Validation and Test

```{r}
set.seed(2019)

val_index <- createDataPartition(
  factor(invoiced[train_index, ]),
  p = 0.2,
  list = F,
  times = 1
)

invoiced_train <- invoiced[train_index, ][-val_index] %>% as.array()
invoiced_val <- invoiced[train_index, ][val_index] %>% as.array()
invoiced_test <- invoiced[-train_index,] %>% as.array()

cat(" Shape of Invoiced Train:", dim(invoiced_train), "\n",
    "Shape of Invoiced Val:", dim(invoiced_val), "\n",
    "Shape of Invoiced Test:", dim(invoiced_test))
```

```{r}
print("--Distribution of Classes in Train--")
print(as.factor(invoiced_train) %>% summary())
print("--Distribution of Classes in Validation--")
print(as.factor(invoiced_val) %>% summary())
print("--Distribution of Classes in Test--")
print(as.factor(invoiced_test) %>% summary())
```

### 1.2 Splitting `Call Text` into Training, Validation and Test

```{r}
call_text_train <- call_text_data[train_index,][-val_index,] %>% as.array()
call_text_val <- call_text_data[train_index,][val_index,] %>% as.array()
call_text_test <- call_text_data[-train_index,]

cat(" Shape of Call Text Train:", dim(call_text_train), "\n",
    "Shape of Call Text Val:", dim(call_text_val), "\n",
    "Shape of Call Text Test:", dim(call_text_test))
```

### 1.2 Splitting `Billing Notes` into Training, Validation and Test

```{r}
billing_notes_train <- billing_notes_data[train_index,][-val_index,] %>% as.array()
billing_notes_val <- billing_notes_data[train_index,][val_index,] %>% as.array()
billing_notes_test <- billing_notes_data[-train_index,]

cat(" Shape of Billing Notes Train:", dim(billing_notes_train), "\n",
    "Shape of Billing Notes Val:", dim(billing_notes_val), "\n",
    "Shape of Billing Notes Test:", dim(billing_notes_test))
```

### 1.3 Splitting `Auxilary` Data into Training, Validation and Test

```{r}
auxillaries_train <- auxillaries[train_index,][-val_index,] %>% as.array()
auxillaries_val <- auxillaries[train_index,][val_index,] %>% as.array()
auxillaries_test <- auxillaries[-train_index,]

cat(" Shape of Auxillaries Train:", dim(auxillaries_train), "\n",
    "Shape of Auxillaries Val:", dim(auxillaries_val), "\n",
    "Shape of Auxillaries Test:", dim(auxillaries_test))
```

## 2. Merging Multiple Inputs

### 2.1 Creating The Input Layers

```{r}
call_text_layer <- layer_input(
  shape = c(CONSTANTS$MAX_LEN),
  name = "call_text_layer"
)

billing_notes_layer <- layer_input(
  shape = c(CONSTANTS$MAX_LEN),
  name = "billing_notes_layer"
)

auxiliary_layer <- layer_input(shape = c(dim(auxillaries)[2]), name = 'auxiliary_layer')
```

### 2.2 Creating The Embedding layers

```{r}
call_text_embedding <- call_text_layer %>%
  layer_embedding(
    input_dim = CONSTANTS$MAX_WORDS,
    output_dim = 512,
    input_length = CONSTANTS$MAX_LEN,
    name = "call_text_embedding") %>%
  layer_dropout(0.6) %>%
  layer_flatten()


billing_notes_embedding <- billing_notes_layer %>%
  layer_embedding(
    input_dim = CONSTANTS$MAX_WORDS,
    output_dim = 512,
    input_length = CONSTANTS$MAX_LEN,
    name = "billing_notes_embedding") %>%
  layer_dropout(0.6) %>%
  layer_flatten()
```

### 2.3 Merging Input and Auxilary Layers

```{r}
main_output <- layer_concatenate(c(call_text_embedding, billing_notes_embedding, auxiliary_layer)) %>%
  layer_dense(units = 64, activation = 'relu', kernel_regularizer = regularizer_l2(l = 0.001)) %>%
  layer_dense(units = 64, activation = 'relu', kernel_regularizer = regularizer_l2(l = 0.001)) %>%
  layer_dense(units = 64, activation = 'relu', kernel_regularizer = regularizer_l2(l = 0.001)) %>%
  layer_dense(units = 1, activation = 'sigmoid', name = 'main_output')
```

### 2.4 Building the Model

```{r}
model <- keras_model(
  inputs = c(call_text_layer, billing_notes_layer, auxiliary_layer),
  outputs = main_output
)
```

```{r}
model %>% compile(
  optimizer = 'rmsprop',
  loss = 'binary_crossentropy',
  metric = c("accuracy")
)

summary(model)
```

```{r plot-model, eval = F}
# plot_model(model, show_shapes = T)
```

### 2.5 Training The Model

```{r echo = T, results = 'hide'}
history <- model %>% fit(
  x = list(call_text_train, billing_notes_train, auxillaries_train),
  y = invoiced_train,
  epochs = 10,
  batch_size = CONSTANTS$BATCH_SIZE_CPU,
  validation_data = list(list(call_text_val, billing_notes_val, auxillaries_val), invoiced_val)
)

# save(history, file = "data/history_model_v5.RData")
# save_model_hdf5(model, filepath = "data/fitted_model_v5")
# save_model_weights_hdf5(model, filepath = "data/fitted_model_v5_weights")
```

```{r}
plot(history,  method = c("ggplot2"))
```

### 2.6 Evaluating The Model

```{r echo = T, results = 'hide'}
model_result <- model %>%
  evaluate(
    list(call_text_test, billing_notes_test, auxillaries_test),
    invoiced_test
  )
```

```{r}
cat('Test loss:', model_result$loss, "\n")
cat('Test accuracy:', model_result$acc, "\n")
```


#### 2.6.1 Model's ROC & AUC

```{r}
pred_probs <- predict(model,
    list(call_text_test, billing_notes_test, auxillaries_test),
    batch_size = CONSTANTS$BATCH_SIZE_CPU
  )

dim(pred_probs)
```

```{r}
roc_result <- roc(invoiced_test, as.vector(pred_probs))

roc_result
```

```{r}
plot(roc_result, col='red', lwd=2)
```

#### 2.6.2 Model's Confusion Matrix

```{r}
pred_class <- as.numeric(pred_probs > .40) %>% as.factor()

confusionMatrix(
  pred_class,
  as.factor(invoiced_test),
  mode = "prec_recall"
)
```