About 40 issues were flagged in the draft. These are the ones that change what the paper can claim. The full list, with locations, is in the error report on the files page.
Class-side comparisonSection 2.2
Announced a comparison of the two directions, but never reported it; the code had no class-side option.
Both directions are implemented and reported on all five datasets over 20 splits.
The paper’s main claim is about direction, so it needs the class-side numbers to stand on.
The α = ∞ collapse on BASEHOCKTable 5
Features 2, 7, 12, 17… listed as the selector's picks, as if they were informative.
Explained as the index walk: every candidate ties at 0, and the argmax returns file order (Proposition 3).
The picks are the file’s column order, so they say nothing about the words.
Choosing kResults
Peak accuracy over the number of features k, read off the test set.
Area under the accuracy curve over all k, with no choice made on test data.
Picking the best k on the test set makes every selector look better than it is.
UncertaintyTables 3–4
A binomial standard error from one split.
20 stratified splits, the Nadeau–Bengio corrected t-test, and Friedman/Nemenyi ranks.
One split cannot separate two selectors; the corrected test accounts for overlapping training sets[7].
SEMEION gainsAbstract, conclusion
Rényi orders α = 2 and 5 reported as beating Shannon.
No order beats Shannon robustly; the gains do not survive 20 splits.
The single-split gains were within noise.
References [12], [14]–[16], [25]Bibliography
JMI, CMIM and mRMR citations permuted; [12] a book review; [25] with the wrong authors.
Yang & Moody for JMI, Fleuret for CMIM, Peng et al. for mRMR, Cover & Thomas for [12], Liu & Hu for [25].
Wrong attributions are the first thing a reviewer checks.
Credit for Proposition 1Section 3
Presented as new.
Credited as the identity-gain case of Theorem 4.2 in Alvim et al.
What is new is what the asymmetry does to selection, not the identity itself[2].