Weird plot of LASSO regression, what's the problem and how to fix it?

I did LASSO regression for my data.

However, the two plots (coefficients plot and cross validation plot) seemed not very well.

The problem with the coefficients plot is: some coefficients grow and then drop, as the lambda changes. In published papers, the coefficients drop as the lambda changes. They don't grow.

coefficients plot

The problem with the cross validation plot is: a part of the red line is not continuous with others.

cross validation plot

My data: https://raw.githubusercontent.com/onkaparinga/default/main/train3.csv

My code:

library(readr)
library(glmnet)

train3 <- read_csv('train3.csv')

x <- as.matrix(train3[,-1])
y <- train3$CustomLabel
cvlasso <- cv.glmnet(x,y,alpha = 1,family = 'binomial')
plot(cvlasso)
plot(cvlasso$glmnet.fit)

Before LASSO regression, I actually did correlation analysis to remove variables that are highly correlated(>0.9) with others.

#read dataset
train2 <- read_csv('train2.csv')

#get non-normalized varibles
non_norm_vars <- train2 %>% 
  summarise(across(1:(ncol(train2)-1),~shapiro.test(.x)$p.value)) %>% 
  t() %>% 
  as.data.frame() %>% 
  filter(V1<0.05) %>% 
  rownames()

#get normalized varibles
norm_vars <- colnames(train2)[!colnames(train2) %in% non_norm_vars] %>% 
  head(-1)

#rearrange the dataset 'train2',put normalized variable in the front, so convinient for replace.
df_nonnorm <- train2[,non_norm_vars]
df_norm <- train2[,norm_vars]
train3 <- bind_cols(df_norm,df_nonnorm)

#calculate coefficient for normalized and overall varibales
cor_norm <- cor(df_norm,method = 'pearson')
cor_all <- cor(train3,method = 'spearman')
#replace the coefficient by normalized variable's
num_norm <- dim(cor_norm)[1]
cor_all[1:num_norm,1:num_norm] <- cor_norm


#sum how many 'high coefficients' (>0.9) each variable has, and rearrange by descendant.
var_seq <- cor_all %>% 
  as_tibble %>% 
  reframe(across(everything(),~sum(abs(.x)>0.9))) %>% 
  t() %>% 
  as.data.frame() %>% 
  arrange(desc(V1)) %>% 
  rownames()
#slice_seq: index of var_seq in colnames(cor_all)
slice_seq <- match(var_seq,colnames(cor_all))
#make a new matrix, most 'high coefficients' variable in the front, least in the end.
cor_all <- cor_all %>% 
  as_tibble() %>% 
  select(all_of(var_seq)) %>% 
  slice(slice_seq) %>% 
  as.matrix()
rownames(cor_all) <- colnames(cor_all)

#set to triangle matrix
cor_all[upper.tri(cor_all)] <- 0
diag(cor_all) <- 0

#keep variables that has 0 'high coefficients'
cor_vars <- cor_all %>% 
  as_tibble() %>% 
  summarise(across(everything(),~any(abs(.x)>0.9))) %>% 
  t() %>% 
  as.data.frame() %>% 
  filter(V1 == F) %>% 
  rownames()
#train3 got
train3 <- train2 %>% 
  select(CustomLabel, all_of(cor_vars))

hope my poor English did not confuse you ...

train2: https://raw.githubusercontent.com/onkaparinga/default/main/train2.csv

Your data set train3 has substantial collinearity, i. e. 4244 bivariate combinations have an r² of at least .9:

corr_mat <- cor(train3[-1])

expand.grid(A = dimnames(corr_mat)[[1]],
            B = dimnames(corr_mat)[[2]]
            ) |>
  cbind(r2 = as.vector(corr_mat)^2) |>
  as_tibble() |>
  filter(as.vector(upper.tri(corr_mat)),
         r2 > .9,
         r2 < 1
         ) |> 
  print(n = 3)

# A tibble: 4,244 x 3
  A     B        r2
  <fct> <fct> <dbl>
1 A693  A1228 0.996
2 A693  A1597 0.988
3 A1228 A1597 0.992
# i 4,241 more rows

Moreover, 31 of your features have at least 3 extreme outliers off the mean by four [sic] standard deviations or more:

train3 |>
summarise(across(where(is.numeric), 
                 ~ sum(abs(mean(.x) - .x) > 4 * sd(.x))
                 )
          ) |> t() |> 
  as.data.frame() |>
  filter(V1 > 3) |>
  nrow()

## + [1] 31

Both multicollinearity and outliers can severely topple your regression.
It would thus be good to put a dimensionality reduction and some outlier management up front of your modelling chain.