This journal includes LIME examples with a logistic regression model fit to the sine data.
library(caret)
library(cowplot)
library(lime)
library(limeaid)
Data
Prepare the sine data
# Prepare training and testing data
x_train = sine_data_train[c("x1", "x2", "x3")]
y_train = factor(sine_data_train$y)
x_test = sine_data_test[1:5, c("x1", "x2", "x3")]
Logistic Regression Models
Main Effects Only
Logistic regression model: y ~ x1 + x2 + x3
# Fit the model
lrm <- caret::train(
as.factor(y) ~ x1 + x2 + x3,
preProcess = "scale",
data = sine_data_train,
method = "glm",
family = "binomial"
)
lrm$finalModel
##
## Call: NULL
##
## Coefficients:
## (Intercept) x1 x2 x3
## -0.1533 -2.2070 1.8365 -0.1240
##
## Degrees of Freedom: 499 Total (i.e. Null); 496 Residual
## Null Deviance: 692.9
## Residual Deviance: 335.8 AIC: 343.8
All Interactions
Logistic regression model: y ~ x1 * x2 * x3
# Fit the model
lri <- caret::train(
as.factor(y) ~ x1 * x2 * x3,
preProcess = "scale",
data = sine_data_train,
method = "glm",
family = "binomial"
)
lri$finalModel
##
## Call: NULL
##
## Coefficients:
## (Intercept) x1 x2 x3 `x1:x2` `x1:x3`
## -0.25910 -2.25052 1.87303 -0.09646 0.24543 0.22155
## `x2:x3` `x1:x2:x3`
## -0.17317 0.01177
##
## Degrees of Freedom: 499 Total (i.e. Null); 492 Residual
## Null Deviance: 692.9
## Residual Deviance: 332.3 AIC: 348.3
Apply LIME
Apply LIME with 3 quantile bins, 4 quantile bins, and kernel density to both models
# Run apply_lime
resm <- apply_lime(
train = x_train,
test = x_test,
model = lrm,
label = "1",
n_features = 2,
sim_method = c('quantile_bins',
'kernel_density'),
nbins = 3:4,
return_perms = TRUE
)
# Run apply_lime
resi <- apply_lime(
train = x_train,
test = x_test,
model = lri,
label = "1",
n_features = 2,
sim_method = c('quantile_bins',
'kernel_density'),
nbins = 3:4,
return_perms = TRUE
)
Exlanation Visualizations
Plot the LIME explanation plot and my EOI plot for each prediction in the test data and each implementation of LIME
# Plot the explanation of interest
join_plots <- function(id, exp) {
p1 <- lime::plot_features(exp$explain[id:(id+1),])
p2 <- limeaid::plot_explain_scatter(exp$explain[id:(id+1),])
plot_grid(p1, p2)
}
Main Effects Only
3 quantile bins
purrr::map(.x = seq(1, 9, 2), .f = join_plots, exp = resm)
## [[1]]
##
## [[2]]
##
## [[3]]
##
## [[4]]
##
## [[5]]
4 quantile bins
purrr::map(.x = seq(11, 19, 2), .f = join_plots, exp = resm)
## [[1]]
##
## [[2]]
##
## [[3]]
##
## [[4]]
##
## [[5]]
kernel density
purrr::map(.x = seq(21, 29, 2), .f = join_plots, exp = resm)
## [[1]]
##
## [[2]]
##
## [[3]]
##
## [[4]]
##
## [[5]]
All Interactions
3 quantile bins
purrr::map(.x = seq(1, 9, 2), .f = join_plots, exp = resi)
## [[1]]
##
## [[2]]
##
## [[3]]
##
## [[4]]
##
## [[5]]
4 quantile bins
purrr::map(.x = seq(11, 19, 2), .f = join_plots, exp = resi)
## [[1]]
##
## [[2]]
##
## [[3]]
