Why Should I Trust You?: Explaining the Predictions of Any Classifier
Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016 · paper · AI and machine learning
Proposes LIME, which explains any classifier's individual predictions by fitting an interpretable model locally, and a method for choosing representative predictions to explain a model. Simulated and human-subject experiments address trust, model selection and improving classifiers.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Reference or label Figure 9 legend labels the second series "Submodular Pick (RP)"; it should be "(SP)", since the first series is "Random Pick (RP)" and the text has "by random (RP) or submodular pick (SP)".
- p. 8 Figure 9 legend
Submodular Pick (RP)
- p. 8 Section 6.2, text under Figure 10
by random (RP) or submodular pick (SP)
Judge's note: The page 8 image shows both legend entries carrying "(RP)". Section 5.1 and Section 6.2 define RP as random pick and SP as submodular pick, and no other meaning of RP is given. The bars and values are unaffected; only the label is wrong.
Second check: another model agreed
Report a problem with this finding
12 minor (typos and the like; most checked by the judge only)
Wording Section 8 says "One issue that we do not mention in this work was how to perform the pick step for images", yet Section 4 does mention it: "For images, I must measure something that is comparable across the super-pixels in different images, such as color histograms or other features of super-pixels; we leave further exploration of these ideas for future work."
- p. 10 Section 8, second paragraph
One issue that we do not mention in this work was how to perform the pick step for
- p. 5 Section 4, end of left column / top of right column
For images, I must measure something
- p. 10 Section 8, second paragraph
One issue that we do not mention in this work was how to perform the pick step for images
- p. 5 Section 4, end of left column / top of right column
For images, I must measure something that is comparable across the super-pixels in different images
Judge's note: Viewed both crops (page 10 conclusion and page 5 Section 4) and searched the text for 'pick step' and 'future work'. The conclusion's wording is as quoted, and Section 4 does raise the pick step for images and defers it to future work. So 'do not mention' is literally inaccurate; 'do not address' was presumably meant. This is a wording inconsistency only, with no effect on data or results.
Second check: another model agreed
Report a problem with this finding
Typo Section 2 prints "offers an alternative method to assess truth in the model"; "truth" should read "trust".
- p. 2 Section 2, second paragraph, left column
offers an alternative method to assess truth in the model
Judge's note: Read on the page 2 image, left column. The preceding sentence ends "thus trust cannot rely solely on it" and the paragraph opens with "measure of overall trust in the model", so this is a word substitution.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.1 prints "We denote x ∈ R^d be the original representation of an instance being explained", mixing "denote ... as" with "let ... be".
- p. 3 Section 3.1, last sentence
denote x ∈ Rd be the original representation
Judge's note: Read on the page 3 image, end of Section 3.1. The sentence is ungrammatical as printed; the meaning is clear.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.2 prints "As not every g ∈ G may be simple enough to be interpretable - thus we let Ω(g) be a measure of complexity"; the "As ... - thus" construction has no proper main clause.
- p. 3 Section 3.2, first paragraph
not every g ∈G may be simple enough to be interpretable -
Judge's note: Read on the page 3 image, Section 3.2 first paragraph. The wording is as quoted; it is a cosmetic syntax slip.
Second check: none (judge only)
Report a problem with this finding
Typo Two grammar slips: "proximity measure between an instance z to x" (should be "and x") and "one of the main purpose of tools like Modeltracker" (should be "purposes").
- p. 3 Section 3.2, second paragraph
proximity measure between an instance z to x
- p. 5 Section 4, third paragraph
one of the main purpose of tools like Modeltracker
Judge's note: Both read on the page images: page 3, Section 3.2 second paragraph, and page 5, Section 4 third paragraph. Both are printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 3.4 prints "cannot be explained by presence of absence of super pixels"; "of absence" should be "or absence".
- p. 4 Section 3.4, last paragraph of left column
cannot be explained by presence of absence of super pixels
- p. 4 Section 3.4, paragraph 'Any choice of interpretable representations'
cannot be explained by presence of absence
- p. 4 Section 3.4, last paragraph
presence of absence of super pixels
Judge's note: Read on the page 4 image, bottom of the left column. Section 3.1 on page 3 uses "presence or absence of a word". The same sentence writes "super pixels" where the paper elsewhere writes "super-pixels".
Second check: none (judge only)
Report a problem with this finding
Typo The procedure is named "K-LASSO in Algorithm 1" in the Section 3.4 text but "K-Lasso(Z, K)" in Algorithm 1.
- p. 4 Section 3.4 vs Algorithm 1
K-LASSO in Algorithm 1
- p. 4 Algorithm 1
K-Lasso(Z, K)
Judge's note: Both forms read on the page 4 image. This is a capitalisation inconsistency only; either form could stand.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 4 caption prints "“Electric Guitar” (p = 0.32), “Acoustic guitar”" while the panel label reads "Electric guitar"; the capitalisation of class names is inconsistent.
- p. 5 Figure 4 caption
“Electric Guitar” (p = 0.32), “Acoustic guitar”
Judge's note: Read on the page 5 image: the caption has "Electric Guitar" beside "Acoustic guitar", and sub-caption (b) has "Electric guitar". Cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Section 5.2 prints "the gold set of features that the are considered important by these models"; there is a stray "the".
- p. 6 Section 5.2, right column below Figure 7
that the are considered important by these models
- p. 6 Section 5.2, first paragraph
features that the are considered important by these models
Judge's note: Read on the page 6 image, right column below Figure 7. The words are continuous in one column, so this is not a column-interleaving artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Section 7 prints "non-interpretable explanations, such as a gradients or linear models with thousands of non-zero weights"; singular article "a" with plural "gradients".
- p. 9 Section 7, last paragraph
such as a gradients or linear
Judge's note: Read on the page 9 image, right column, last paragraph. Printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.4 closes with "for getting insights into classifiers knowing when not to trust them and why"; a comma or "and" is missing between "classifiers" and "knowing".
- p. 9 Section 6.4, last sentence
getting insights into classifiers knowing
- p. 9 Section 6.4, last sentence
getting insights into classifiers knowing when not to trust them and why
Judge's note: Read on the page 9 image, end of Section 6.4. As printed, the sentence reads as if the classifiers do the knowing. A missing conjunction or punctuation mark is the evident cause.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [22] prints "AAAI Spring Syposium on Knowledge Representation and Reasoning"; "Syposium" should be "Symposium".
- p. 10 References, entry [22]
AAAI Spring Syposium on Knowledge
Judge's note: Read in reference [22] on the page 10 image, and the text layer has the same spelling ("Syposium"). It is a plain misspelling.
Second check: none (judge only)