Plot - shapr

[hanneleer]

Dear all,

I attempted to plot the Shapley values using the shapr package, but I encountered an issue. Here is the plot I generated:

Has anyone else experienced a similar issue? I don’t think the plot is displaying correctly, especially with the strange vertical lines. Any advice would be greatly appreciated!

Thanks!

[martinju]

Hi. I agree this does not look good. I thought we had fixed things like this in #406, but maybe this is an edge case. Please provide a complete runnable example, and we will look into it.

[hanneleer]

Hi @martinju! Thank you so much for this! I hope this is enough information - please let me now otherwise:

Data_O <- read_csv(synthetic_data)
# Remove rows with missing values
Data_O <- Data_O[complete.cases(Data_O),]
# Handle extremes of target 
Data_O <- Data_O %>% filter(actief_in_inst_2022_SCH > 0.60)
Data_O$actief_in_inst_2022_SCH <- sqrt(Data_O$actief_in_inst_2022_SCH)

# Features 
check <- as.data.frame(model.matrix(~., data = Data_O[, c(3, 32:36, 38, 55:68)]))
check[] <- lapply(check, as.numeric)  
check <- as.matrix(check)  
check <- check[, -1]  

# Outcome variable
y <- as.numeric(Data_O$actief_in_inst_2022_SCH)

# Split dataset into training (70%) and test (30%) sets
samp <- sample(nrow(Data_O), 0.7 * nrow(Data_O))

Train1 <- check[samp, ]
Train1 <- as.data.frame(Train1)

Test1 <- check[-samp, ]
Test1 <- as.data.frame(Test1)

Y_train <- y[samp]
Y_test <- y[-samp]

# Train Random Forest model 
rf.fit <- ranger::ranger(Y_train ~ .,
                         data = Train1,
                         mtry = 14,  
                         max.depth = 3,
                         replace = FALSE, 
                         min.node.size = 40, 
                         sample.fraction = 0.8, 
                         respect.unordered.factors = "order", 
                         importance = "permutation")

# SHAPR 
p <- mean(Y_train)
library(shapr)
explanation <- shapr::explain(
  rf.fit,
  Test1, 
  Train1,
  approach = "gaussian",
  phi0 = p
)

library(ggplot2)
library(ggbeeswarm)
# Plot 
if (requireNamespace("ggplot2", quietly = TRUE)) {
  plot(explanation, plot_type = "scatter")
  plot(explanation, plot_type = "beeswarm")
}


[synthetic_data.csv](https://github.com/user-attachments/files/17777841/synthetic_data.csv)


Thanks! 
Hanneleer 

[hanneleer]

Hi @martinju. I was wondering if you have had a chance to look into the issue. I’m still struggling to figure out what might be going wrong and whether there’s something I might be misunderstanding or doing incorrectly. I really appreciate any insights you can share whenever you have time!

[martinju]

Hi. Just confirming that I have started looking into this. I have fixed the vertical issue by using corral for scaling in ggbeeswarm instead, to then realizing the original issue might be the horizontal scaling. Not sure what is going on there. Will look more into it tomorrow/Friday.

[martinju]

Hi again

#424 should fix the issue if you use plot(explanation, plot_type = "beeswarm", corral = "wrap")

Please confirm that it fixes the issue for you. You can now also further control the behavior of the beeswarm plot with the ... arguements passed to ggbeeswarm::geom_beeswarm().

[hanneleer]

Thanks a lot @martinju !! It works, really appreciate it!!