Changing the num_features parameter in explain_instance gives very different results

I have recently started using LIME and I’m using it to explain the outputs of a LGBMClassifier.

I have an issue with the num_features parameter. For example if I run it with 5, the output is something like this:

exp = explainer.explain_instance(X_train.values[idx], predict_model_lgbm.predict_proba, num_features=5) exp.as_list()

[(‘entitlement_number <= 1.00’, -0.41398972465855205), (‘Annual=0’, 0.33586503353921526), (‘Daily=0’, -0.23173429118453642), (‘hist_total_tenurea <= 0.00’, 0.210909651734952), (‘numrenewals > 1.00’, 0.12207210974714447)]

But if I run it with num_features = 10, I get something like this:

[(‘SAILING <= 0.00’, 0.0), (‘LIVE ONLY <= 1.00’, 0.0), (‘MODERN PENTATHLON <= 0.00’, 0.0), (‘JET SKIING <= 0.00’, 0.0), (‘VOLLEYBALL <= 0.00’, 0.0), (‘HANDBALL <= 0.00’, 0.0), (‘SHOW JUMPING <= 0.00’, 0.0), (‘One Time=0’, 0.0), (‘ARTISTIC GYMNASTICS <= 0.00’, 0.0), (‘XTREME SPORTS <= 0.00’, 0.0)]

I would have thought that increasing the features from 5 to 10 would keep the original 5 plus adding another 5 that are less important. Am I using this parameter correctly?

On another note, the features shown by the explainer are very different to the ones I get when plotting model’s feature importances (using plot_importance from Lightgbm). Is that something I should expect?

[pancodia]

Any update on this issue?

[sandymoens]

Hi Marco, first of, thanks for the implementation 👍. Secondly, I am using lime==0.2.0.1 and am getting a similar problem as Fernando. For num_features < 7 the explanations are generally incremental and intuitive. Using num_features between 7 and 20 i get zeros on explanations and going even higher, I get either zeros or unintuitive features explanations, i.e., explanations altering from one class to another and values larger than the first 6 “good” explanations. The dataset is non public so unfortunately i cannot pass a copy for reproduction purposes.