An experiment on autoencoding Wisconsin Breast Cancer Diagnosis dataset

wdbc.rmd

---
title: "WDBC"
author: "David Charte"
date: "22 de noviembre de 2017"
output:
  html_document: default
---

```{r setup, include=FALSE, echo = F, results = "hide", message=FALSE}
knitr::opts_chunk$set(echo = FALSE)
library(foreign)
library(RWeka)
library(xgboost)
library(ruta)
library(rutavis)
library(e1071)
library(partykit)
library(randomForest)
library(caret)
library(webshot)

# wdbc <- read.arff("wdbc.arff")
wdbc <- read.csv("wdbc.csv")

cv <- function(data, predictions) {
  set.seed(12345678)
  
  folds <- createFolds(data[, 1], k = 10)

  accuracy <- sapply(folds, function(test) {
    pred <- predictions(data[-test, ], data[test, -1])
    sum(pred == data[test, 1]) / length(test)
  })
}

model <- function(method) {
  list(
    c45 = function(tra, tst) {
      predict(J48(class ~ ., data = tra), tst)
    },
    svm = function(tra, tst) {
      predict(svm(class ~ ., data = tra), tst)
    },
    rf  = function(tra, tst) {
      predict(randomForest(class ~ ., data = tra), tst)
    },
    xg  = function(tra, tst) {
      mod <- xgboost(data = as.matrix(tra[, -1]), 
                    label = as.numeric(tra[, 1]) - 1, 
                    max_depth = 2, eta = 1, nthread = 4, 
                    nrounds = 2, objective = "binary:logistic")
      pred <- predict(mod, as.matrix(tst))
      pred <- as.numeric(pred > 0.5)
      pred <- as.factor(pred)
      levels(pred) <- levels(tra[, 1])
      pred
    }
  )[[method]]
}

autoencode <- function(dat, cl = 1, hidden, seed = 8912345, ...) {
  set.seed(seed)
  mxnet::mx.set.seed(seed)
  
  ae <- ruta.makeLearner("autoencoder", 
                         hidden = c(ncol(dat) - 1, hidden, ncol(dat) - 1), 
                         activation = "leaky")
  
  tk <- ruta.makeTask(data = dat, cl = cl)
  
  aemd <- ruta::train(
    ae,
    tk,
    epochs = 200,
    optimizer = "adam",
    #momentum = 0.002,
    learning.rate = 0.02,
    initializer.scale = 1,
    wd = 0.01,
    ...
  )
  
  encoded <- as.data.frame(ruta.deepFeatures(aemd, tk))
  encoded <- cbind(dat[, 1], encoded)
  names(encoded) <- c("class", paste0("v", as.character(1:(ncol(encoded)-1))))
  list(aemd, encoded, tk)
}


run <- function(methods) {
  res <- ""
  encoded5 <- autoencode(wdbc, hidden = 5)
  encoded3 <- autoencode(wdbc, hidden = 3)
  
  for (method in methods) {
    res <- paste0(res, "Method: ", method, "\n==================\n")
    ac <- cv(wdbc, model(method))
    res <- paste0(res, "Model trained with original dataset:\n")
    res <- paste0(res, "  Mean accuracy: ", mean(ac), ", std dev:", sd(ac))
    
    ac <- cv(encoded5[[2]], model(method))
    res <- paste0(res, "\nModel trained with 5 autoencoded variables:\n")
    res <- paste0(res, "  Mean accuracy: ", mean(ac), ", std dev:", sd(ac))
    
    ac <- cv(encoded3[[2]], model(method))
    res <- paste0(res, "\nModel trained with 3 autoencoded variables:\n")
    res <- paste0(res, "  Mean accuracy: ", mean(ac), ", std dev:", sd(ac))
    res <- paste0(res, "\n\n")
  }
  
  list(res, encoded5[[1]], encoded3[[1]], encoded5[[3]])
}
```

### Accuracy en test de los algoritmos

```{r, echo = F, results = "hide", message=FALSE, warning = F}
res <- run(c("c45", "svm", "rf", "xg"))
```

```{r, echo = F}
cat("WDBC is a binary dataset consisting of", nrow(wdbc), "observations of", ncol(wdbc), "features.")
```

```{r, echo = F}
cat(res[[1]])
```

```{r, echo = F, warning = F, message = F}
plot(res[[3]], res[[4]], size = 1, sizes = c(1, 2))
```

## Árboles de decisión

```{r, echo = F, results="hide", message=F, warning=F}
set.seed(12345678)
encoded5 <- autoencode(wdbc, hidden = 5)[[2]]
encoded3 <- autoencode(wdbc, hidden = 3)[[2]]

trees <- lapply(list(wdbc, encoded5, encoded3), function(df) {
  print(dim(df))
  folds <- createFolds(df[, 1], k = 10)
  J48(class ~ ., data = df[-folds[[1]],])
})
```

### Datos originales

```{r}
print(trees[[1]])
```

### Datos del autoencoder de 5 variables

```{r}
print(trees[[2]])
```

### Datos del autoencoder de 3 variables

```{r}
print(trees[[3]])
```