LimeTextExplainer returns a lot zeros only explanations

OS: Manjaro x64 Python: 3.6.11 Lime: 0.1.1.32

I am using LimeTextExplainer to explain instances of different models (Logistic Regression, SVM, XGBoost, etc.). Explaining round about 800 instances always results in 300-400 explanations that contain zeros only.

For example: Instance: [‘This’, ‘ball’, ‘is’, ‘red’, ‘and’, ‘gold’] Explanation: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

I thought it might be the length of the data instance but it’s not. There is no pattern on first view. I have very short instances with “real” explanations and very short with only zeros. Same for long ones.

Code:

explainer = LimeTextExplainer()
exp = explainer.explain_instance(text, clf.predict, num_features=len(text.split()), num_samples=1000)

Any idea what I miss?

[marcotcr]

Since you’re using default parameters, LIME will remove words at random. Is it the case that the prediction of your model is constant for all variations of ‘This ball is red and gold’? i.e. what output do you get if you run the following?

clf.predict(['This ball', 'This', 'This ball is red', 'This ball is red gold', 'This ball is red and gold'])

I suggest using predict_proba to get a bit of signal from the prediction probability

[marcotcr]

It could be that your model is not changing prediction probabilities much around those perturbations and / or that the explanation is just the intercept (see #304, for example).