Reproduction of the draft
An independent implementation reproduced every first-pick relevance, selection, score and accuracy in the draft’s Tables 1–5 on the stratified seed-0 split (for example BASEHOCK \(\alpha=\infty\): feature 1721, 0.1379; SEMEION \(\alpha=5\): feature 161, 0.605). The reported numbers are feature-side \(I_\alpha(f;C)\), as the text says.
Errors in the first draft
About 40 issues were flagged, each with its location and fix, in the error report. The ones that change conclusions are below; What changed shows each as before and after.
| Issue | Why it matters |
|---|---|
| Abstract unfinished | Must be written before submission. |
| Class-side comparison announced (Section 2.2) but never reported; the code has no class-side option | The conclusion about direction was unsupported. |
| \(\alpha=\infty\) collapse on BASEHOCK | A provable degeneracy: Table 5’s features 2, 7, 12, 17… are the file’s column order. |
| Peak accuracy over \(k\) taken on the test set | Optimistically biased. |
| Binomial standard error on one split | Wrong uncertainty for comparing two selectors. |
| SEMEION gains at \(\alpha=2, 5\) | Do not hold up over 20 splits. |
| References [14]–[16] permuted | JMI is Yang & Moody, CMIM is Fleuret, mRMR is Peng et al. |
| [12] cites a book review | Should be Cover & Thomas. |
| Palamidessi–Romanelli objection to Endo–Kudo | Stated backwards. |
The first review called reference [25] unverifiable. It exists: C.-T. Liu & B.-G. Hu, “Mutual information based on Renyi’s entropy feature selection”, IEEE ICIS 2009. The draft had the authors wrong (“X. Xu et al.”).
Originality check
No sign of copied text. Two results needed credit: Proposition 1 is the identity-gain case of Theorem 4.2 in Alvim et al. (CSF 2012)[2], and Palamidessi & Romanelli already note that \(I_\infty\) is not symmetric[1]. What is new is what the asymmetry does to selection.
Experimental results
| Comparison | Result |
|---|---|
| BASEHOCK AUC at \(\alpha=\infty\), feature-side → class-side | 57.9% → 89.7% |
| Text sets: feature-side \(\alpha=\infty\) vs Shannon | −22 to −33 pts |
| Text sets: class-side \(\alpha=\infty\) vs Shannon | −1.4 to −5.9 pts |
| SEMEION, GISETTE: gap between directions at \(\alpha=\infty\) | ≤ 1.6 pts |
The class-side blind zone
With unequal class sizes, \(I_\infty(C;f)\) is zero whenever the most likely class is the same with and without the word. For two classes, \(I_\infty(C;f) = \log_2\!\big(\max\{\pi_1-\delta,\ \pi_1,\ \pi_2+\delta\}/\pi_1\big)\), which is zero exactly when \(0 \le \delta \le \pi_1-\pi_2\). This, not estimation noise, explains the synthetic class-side drop and the poor class-side results on RELATHE (classes 779 vs 648).
Second classifier and joint rule
Bernoulli naive Bayes was added, because Peng et al. (mRMR)[13] and Fleuret (CMIM)[15] both evaluated with naive Bayes. The Palamidessi–Romanelli joint-conditioning rule has the best average rank of the eleven selectors with both classifiers, and beats Shannon mRMR by 1.7–3.5 points on PCMAC and RELATHE.