library(tidyverse)

df = read_csv("creditcard.csv")
df = as.data.frame(df)
dffraud = df[df$Class==1, ]
dffraud = sample_n(dffraud, 400)
dfnonfraud = df[df$Class==0, ]
dfnonfraud = sample_n(dfnonfraud, 50000)

dftrain = rbind(dffraud, dfnonfraud)
dftrain = as.data.frame(dftrain)
y = dftrain$Class == 1
sigmoid <- function(x) 1 / (1 + exp(-x))

feedforward <- function(x, w1, w2) {
  z1 <- cbind(1, x) %*% w1
  h <- sigmoid(z1)
  z2 <- cbind(1, h) %*% w2
  list(output = sigmoid(z2), h = h)
}

backpropagate <- function(x, y, y_hat, w1, w2, h, learn_rate) {
  dw2 <- t(cbind(1, h)) %*% (y_hat - y)
  dh  <- (y_hat - y) %*% t(w2[-1, , drop = FALSE])
  dw1 <- t(cbind(1, x)) %*% (h * (1 - h) * dh)

  w1 <- w1 - learn_rate * dw1
  w2 <- w2 - learn_rate * dw2

  list(w1 = w1, w2 = w2)
}

train <- function(x, y, hidden = 5, learn_rate = 1e-3, iterations = 1e5) {
  d <- ncol(x) + 1
  w1 <- matrix(rnorm(d * hidden), d, hidden)
  w2 <- as.matrix(rnorm(hidden + 1))
  for (i in 1:iterations) {
    ff <- feedforward(x, w1, w2)
    bp <- backpropagate(x, y,
                        y_hat = ff$output,
                        w1, w2,
                        h = ff$h,
                        learn_rate = learn_rate)
    w1 <- bp$w1; w2 <- bp$w2
  }
  list(output = ff$output, w1 = w1, w2 = w2)
}

x = dftrain[, -dftrain$Class]
x = as.matrix(x)

#test

#nnet = feedforward(x, nnet5$w1, nnet5$w2)
#mean((nnet$output > .5) == y)
#frauds = length(which(nnet$output > 0.05))
#nonfrauds = length(which(nnet$output < 0.05))

#train NN
#nnet5 <- train(x, y, hidden = 300, iterations = 1e5)
#mean((nnet5$output > .5) == y)

save(nnet5, file = "nnet300_400_50000.RData")