Deep learning, neural network - r

Question

I have a question regarding applying a neural network in categorical data.
1- I have one output which is numeric (Connection.Duration)
2- I have 5 inputs, 4 of them (EVSE.ID, User.ID, Fee, Day) are categorical and 1 (Time) is numeric.
I want to apply a neural network to predict the Connection.Duration. I do not know the correct command to use for categorical data. I used model.matrix but I did not how to continue with the new data frame (m) which contains the categorical data.
I would like to ask for help please.
data$Fee <- as.factor(data$Fee)
data$EVSE.ID <- as.factor(data$EVSE.ID)
data$User.ID <- as.factor(data$User.ID)
data$Day <- as.factor(data$Day)
data$Time <- as.factor(data$Time)
data$Connection.Duration <- as.factor(data$Connection.Duration)
m <- model.matrix(Connection.Duration ~ EVSE.ID+Time+Day+Fee+User.ID,
data= data)
# Neural Networks
n <- neuralnet(Connection.Duration ~ EVSE.ID+Time+Day+Fee+User.ID,
data = m,
hidden=c(100,60))
# Data partition
set.seed(1234)
ind <- sample(2, nrow(m), replace = TRUE, prob = c(0.7, 0.3))
training <- m[ind==1,1:5]
testing <- m[ind==2,1:5]
trainingtarget <- m[ind==1, 6]
testingtarget <- m[ind==2, 6]
# Normalize
m <- colMeans(training)
s <- apply(training, 2, sd)
training <- scale(training, center = m, scale = s)
testing <- scale(testing, center = m, scale = s)
# Create Model
model <- keras_model_sequential()
model %>%
layer_dense(units = 5, activation = 'relu', input_shape = c(5)) %>%
layer_dense(units = 1)
# Compile
model %>% compile(loss= 'mse',
optimizer= 'rmsprop',
metrics='mae')
# Fit model
mymodel <- model %>%
fit(training,
trainingtarget,
epochs= 100,
batch_size = 32,
validation_split = 0.2)
# Evaluate
model %>% evaluate(testing, testingtarget)
pred <- model %>% predict(testing)
mean(testingtarget- pred^2)
plot(testingtarget, pred)
# Fine-tune Model
model <- keras_model_sequential()
model %>%
layer_dense(units = 100, activation = 'relu', input_shape = c(5)) %>%
layer_dropout(rate = 0.4) %>%
layer_dense(units = 60, activation = 'relu', input_shape = c(5)) %>%
layer_dropout(rate = 0.2) %>%
layer_dense(units = 1)
# Compile
model %>% compile(loss= 'mse',
optimizer= optimizer_rmsprop(lr=0.0001),
metrics='mae')
# Fit model
mymodel <- model %>%
fit(training,
trainingtarget,
epochs= 100,
batch_size = 32,
validation_split = 0.2)
# Evaluate
model %>% evaluate(testing, testingtarget)
pred <- model %>% predict(testing)
mean(testingtarget- pred^2)
plot(testingtarget, pred)

Answer 1

What you're looking for is called "one hot encoding". There are functions in tensorflow/keras to help out with the encoding.
But otherwise, I would try to do it up front. I would not rely on model.matrix as it doesn't give you quite what you want.
You can easily write your own function, but here's an example using the mltools package:
library(data.table)
library(mltools)
one_hot(data.table(x = factor(letters), n = 1:26))
Note: it requires data.table rather than data.frame but you can convert your data back and forth.