I am trying to understand why changing the reference level of a factor changes the results of a model. Consider this example:

library(liver)
library(caret)
library(glmnet)
library(dplyr)

data(churn)
head(churn)

# set reference levels
churn$state <- relevel(churn$state, ref = "NE")
churn$area.code <- relevel(churn$area.code, ref = "area_code_408")
churn$intl.plan <- relevel(churn$intl.plan, ref = "yes")
churn$voice.plan <- relevel(churn$voice.plan, ref = "no")

# split into train and test
set.seed(1)
train.index <- createDataPartition(churn$churn, p = 0.8, list = FALSE)
train_churn <- churn[train.index,]
test_churn  <- churn[-train.index,] 

# add class weights
my_weights = train_churn %>%
  select(churn) %>%
  group_by(churn) %>%
  count()

weight_for_yes = (1 / my_weights$n[1]) * ((my_weights$n[1] + my_weights$n[2]) / 2.0)
weight_for_yes
weight_for_no = (1 / my_weights$n[2]) * ((my_weights$n[1] + my_weights$n[2]) / 2.0)
weight_for_no

model_weights <- ifelse(train_churn$churn == "yes", weight_for_yes, weight_for_no)

# tuning grid
myGrid <- expand.grid(
  alpha = 0, 
  lambda = seq(0,1,0.01)  
)

set.seed(1)
mod_1 <- train(churn ~
                 state +
                 area.code +
                 intl.plan +
                 voice.plan,
               data = train_churn,
               method = "glmnet",
               tuneGrid = myGrid,
               weights = model_weights)

mod_1

Tuning parameter ‘alpha’ was held constant at a value of 0
Accuracy was used to select the optimal model using the largest value.
The final values used for the model were alpha = 0 and lambda = 0.8

prediction <- predict(mod_1, newdata = test_churn)
confusionMatrix(prediction, test_churn$churn)

I also look at a single new prediction

new_data = data.frame(state = c("CA"),
                      area.code = c("area_code_510"),
                      intl.plan = c("yes"),
                      voice.plan = c("no"))

predict(mod_1, newdata = new_data, type = "prob")

Now I restart R, set new reference levels, and rerun all the code. This is the output

# set new reference levels
churn$state <- relevel(churn$state, ref = "OR")
churn$area.code <- relevel(churn$area.code, ref = "area_code_415")
churn$intl.plan <- relevel(churn$intl.plan, ref = "no")
churn$voice.plan <- relevel(churn$voice.plan, ref = "yes")

Tuning parameter ‘alpha’ was held constant at a value of 0
Accuracy was used to select the optimal model using the largest value.
The final values used for the model were alpha = 0 and lambda = 0.62.

I expected the lambdas to change, but I did not expect the confusion matrix or classification probabilities to changes. Is this normal?