week2.Rmd

---
title: "H~2~O framework introduction for predictive analysis"
output: html_document
---

## Exercise: Comparing Models
In this exercise our objective will be building a model to identify customers most likely to accept a direct marketing offer.  

I'll be using H~2~O flow, which runs on the browser and is designed to streamline the modelling process. It offers the same functionaliity as the non GUI H~2~O, as well as more insightful output. I'll be using the dataset I imported in the last step, but would import it again here for reproducibility reasons.  

```{r include = FALSE, echo=FALSE}
library(downloader)
library(h2o)
library(tidyverse)
```

```{r warning=FALSE}
localH20 <- h2o.init(ip = "127.0.0.1", port = 54321)
```

The above starts the h~2~o server on the ip given address. That is where we import our data, prepare for modelling, and draw our model.  

In the previous exercise on H~2~O Flow, I created a neural net and GBM model, which I compared for performance and predicted the *y* column. In this next exercise which I'll be doing here with H~2~O as well, I'll be creating a classifier to know if a customer would accept a beneficial offer, or decline. I'll be using a Logistic Regression model for this classifier. Also, I'll be predicting how long a call would last.


### Upload file, and explore   

```{r warning=FALSE, include=FALSE}
# dataUrl = "https://raw.githubusercontent.com/QUT-BDA-MOOC/FLbigdataStats/master/bank_customer_data.csv"
# download.file(url <- dataUrl, destfile = "BankCustomerData.csv")
user_data <- h2o.uploadFile("BankCustomerData.csv", destination_frame = "", parse = T, header = T,
                                   sep = ",", na.strings = c("unknown"), progressBar = FALSE,
                                   parse_type = "CSV" )
```

``` {r warning=FALSE}
summary(user_data)
class(user_data)
head(user_data)
table(as.data.frame(user_data)$y)
h2o.describe(user_data)

# Gets the index of a column, takes the columnNames, and the column to find the index.
getColumnIndex <- function(frameColumnNames, column) {
  index = frameColumnNames %in% column
  if (any(index)) {
    return(which(index))
  } else {
    return(NULL)
  }
}

# Exclude the length column from the data set
user_data_predict_offer_accept <- user_data[, -11] # excluding call length
(split_frame_offer <- h2o.splitFrame(user_data_predict_offer_accept, ratios = 0.75))
train_frame_offer <- split_frame_offer[[1]]
test_frame_offer <- split_frame_offer[[2]]
(frame_column_names_offer <- colnames(train_frame_offer))
outcomeIndex = getColumnIndex(frame_column_names_offer, "y")
predictorsIndices <- (1:length(frame_column_names_offer))[-outcomeIndex]

offer_glm <- h2o.glm(x=predictorsIndices, y = outcomeIndex, training_frame = train_frame_offer, 
                     validation_frame = test_frame_offer, max_iterations = 100, solver = "L_BFGS", 
                     family = "binomial", alpha = 1, intercept = T)
                      

user_data_predict_call_length <- user_data[, -21] # excluding offer accept
(split_frame_call <- h2o.splitFrame(user_data_predict_call_length, ratios = 0.75))
train_frame_call <- split_frame_call[[1]]
test_frame_call <- split_frame_call[[2]]
(frame_column_names_call <- colnames(train_frame_call))
(outcomeIndex = getColumnIndex(frame_column_names_call, "length"))
(predictorsIndices <- (1:length(frame_column_names_call))[-outcomeIndex])

# could normalized the values before this normalizing
call_glm <- h2o.glm(x=predictorsIndices, y = outcomeIndex, training_frame = train_frame_call, 
                    validation_frame = test_frame_call,max_iterations = 100, solver = "L_BFGS", 
                    family = "gaussian", alpha = 1,intercept = T)

summary(offer_glm)
summary(call_glm)

# making the predictions, should use an appropriate data set, not this ones used for validation in training
h2o.predict(call_glm, test_frame_call)
h2o.predict(offer_glm, test_frame_offer)

```