Work

One Dataset Too Many: The Fifty-Fourth Row in the Benchmark Tables of Bertsimas and Dunn

Year: 2026
Type: preprint
Abstract: Bertsimas and Dunn report in Machine Learning (2017) optimal classification trees, OCT and OCT-H, benchmarked against CART on what the paper describes throughout, in its abstract and in Sections 1 and 5, as 53 UCI datasets. Every full-results table in the paper prints 54 datasets. The additional row is Optical-recognition, and it is a row in name only: every summary aggregate the paper reports, in Tables 6, 8, and 12, reproduces exactly, to printed precision, when Optical-recognition is excluded, and none of them reproduces when all 54 printed rows are used. The discrepancy is consistent across both methods and all four depths, and Optical-recognition is the unique single row whose removal restores agreement. The paper's text and summaries therefore describe a 53-dataset benchmark while its results tables print a 54th dataset whose numbers enter no reported aggregate. Either the row or the summaries need correction.
Language: English
Open Access status: green
OpenAlex ID: W7228340813
Cites:
Cited by:
Related to: 10