My question is related to this one: LIME explanation confusion. But since it does not have a reproducible example or an answer, I am asking here with an example.

I have a dataset with unbalanced classes. Here I make a reproducible example with the wine data from the breakDown package. My actual data is even more unbalanced than this example. I train a gbm model with the caret package. Below I produce plots with the lime package for two correct predictions and two incorrect predictions.

My question is about the lime explanation plots and how the probabilities of the predictions relate to the weights. Here are the plots for the correct cases: enter image description here

As you can see, case 1364 has a probability of 0.779 for the zero case, and it is indeed classified correctly as a zero. However, from my understanding of the lime results, the data from this case do not support this class (the red/negative weights). My question is, how can the model predict this high probability for the zero class, and the lime explanation tell us it is not supported? Contrast this to case 166, where it appears there is support for high weight variables for the zero probability class and this case also has a high probability for this class (0.981).

Does it have to do with my dataset having unbalanced classes? I tried lowering the threshold for classification into a one class, and I still see many cases where the lime plots seem to contradict the probabilities.

Example below:


# some code from: https://shiring.github.io/machine_learning/2017/04/23/lime

# make an unbalanced classifier for the wine data quality
new.wine <- transform(wine, sdclass = cut(quality, 
                                          breaks=c(-Inf, 6, Inf), labels=c("zero", "one")))

# remove the quality column
new.wine <- new.wine[,-12]

# create a training and a testing dataset
new.wine$spl=sample.split(new.wine$sdclass, SplitRatio = 0.7)

train <- subset(new.wine, new.wine$spl == TRUE)[,-13]
test <- subset(new.wine, new.wine$spl == FALSE)[,-13]


# train a gbm model
trainControl <- trainControl(method="cv", number=10)

model <- train(sdclass~., data = train, 
               verbose = FALSE)


# predictions from caret gbm model
caret_pred <- predict.train(model, newdata = test, type="prob") 
obs <- ifelse(test$sdclass == "zero" , 0, 1)
SDMTools::accuracy(obs, caret_pred[[2]], threshold = 0.5)
#threshold       AUC omission.rate sensitivity specificity prop.correct     Kappa
#1       0.5 0.6862422     0.5597484   0.4402516   0.9322328    0.8257318 0.4203127

# make a lime model explainer
explainer <- lime(x = train, model = model, bin_continuous = TRUE, n_bins = 5)

# get predictions for the model 
pred <- data.frame(sample_id = 1:nrow(test),
                   actual = test$sdclass)
pred$prediction <- colnames(pred)[2:3][apply(pred[, 2:3], 1, which.max)]
pred$correct <- ifelse(pred$actual == pred$prediction, "correct", "wrong")

# filter the correct and wrong predictions
pred_cor <- filter(pred, correct == "correct")

pred_wrong <- filter(pred, correct == "wrong") 

# select two test samples with correct predictions
test_data_cor <- test %>%
  mutate(sample_id = 1:nrow(test)) %>%
  filter(sample_id %in% pred_cor$sample_id) %>%
  sample_n(size = 2) %>%
  remove_rownames() %>%
  tibble::column_to_rownames(var = "sample_id") %>%

# select two samples with wrong predictions
test_data_wrong <- test %>%
  mutate(sample_id = 1:nrow(test)) %>%
  filter(sample_id %in% pred_wrong$sample_id) %>%
  sample_n(size = 2) %>%
  remove_rownames() %>%
  tibble::column_to_rownames(var = "sample_id") %>%

# explain selected test data
explanation_cor <- lime::explain(x = test_data_cor, explainer, n_features = 4, n_labels = 2)
explanation_wrong <- lime::explain(x = test_data_wrong, explainer, n_features = 4, n_labels = 2)

# plot the explanation
plot_features(explanation_cor, ncol = 2)
plot_features(explanation_wrong, ncol = 2)

1 Answer 1


What LIME does is fitting a "simple and interpretable" model on top of "complex and uninterpretable" model. That simple model is assumed to be a good approximation of the complex model's behaviour in the vicinity of the case we want to examine. (To be a bit more exact: The data points used to fit the simple model are weighted based on the similarity to the case $i$ we want to explain, points very similar to case $i$ have higher weights.) If the approximation provided by the simple model is bad, then our explanation is unreliable. This is exactly what happens here.

By default, LIME uses a generalised linear model as provided by the function glmnet with $\alpha =0$, i.e. a standard linear model with ridge penalty. If we check the $R^2$ values of the ridge regression models used here, in all cases the values are between $0.069$ and $0.075$. This means that the simple model explanations capture a rather small proportion of the variance of the complex model. For that reason, interpreting these coefficients is marginally useless. LIME is not a good tool for this use case. LIME is a really great tool but it is not a silver bullet when it comes to ML interpretation.

  • $\begingroup$ In the models I have tried, the R2 is usually around 0.1 to 0.2. At first I figured my model was highly complex as we are modeling an environmental system with many variables. However, the R2 is even lower for this example data with few variables. I have only seen the very simple model examples on very small data where LIME had reasonable R2, for example the iris dataset. Have you had luck applying it to your real data with ml algorithms? $\endgroup$
    – user29609
    Jul 3, 2018 at 21:32
  • $\begingroup$ Yes, I have. I have seen some ~0.40 with real data; never 0.50+. It also depends where the particular point is within the sample space. Some points are just harder (or easier) to fit locally. Check the package iml and in particular the function Shapley it is an excellent alternative to LIME. $\endgroup$
    – usεr11852
    Jul 3, 2018 at 22:38

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.