MLbias

MLbias corrects selection bias when comparing machine-learning classifiers by fitting learning curves with an inverse power law to adjust cross-validation error rates for more reliable genomic classifier evaluation.


Key Features:

  • Implementation: Implemented as an R package for analysis and correction of classifier error-rate estimates.
  • Bias correction methodology: Uses learning curve fitting by inverse power law (IPL) to mitigate selection bias from choosing the classifier with the smallest cross-validation error rate.
  • Comparative methods: Explicitly compares IPL-based correction against nested cross-validation, weighted mean correction, and the Tibshirani-Tibshirani procedure.
  • Performance characterization: Demonstrated via extensive simulations and real-world dataset analyses that IPL yields superior bias correction with smaller variance.
  • Error extrapolation: Extrapolates error estimates to larger sample sizes to inform whether additional samples would improve classifier accuracy.
  • Targeted domain: Intended for genomic studies with moderate sample sizes (e.g., n = 30–60) where reported cross-validation error rates may be optimistic.

Scientific Applications:

  • Genomic classifier development: Adjusts cross-validation error rates to support development of classifiers that generalize to independent testing datasets.
  • Clinical study validation: Reduces the risk of reporting artificially deflated error rates in clinical studies by correcting selection bias.
  • Sample size recommendations: Provides guidance on whether additional data collection is warranted by extrapolating expected error at larger sample sizes.

Methodology:

Fits learning curves using an inverse power law (IPL), compares IPL-corrected error estimates with nested cross-validation, weighted mean correction, and the Tibshirani-Tibshirani procedure, evaluates performance via simulations and real-world dataset analyses, and extrapolates error estimates to larger sample sizes.

Topics

Details

Tool Type:
command-line tool
Operating Systems:
Linux, Windows, Mac
Programming Languages:
R
Added:
8/3/2017
Last Updated:
11/25/2024

Operations

Publications

Ding Y, Tang S, Liao SG, Jia J, Oesterreich S, Lin Y, Tseng GC. Bias correction for selecting the minimal-error classifier from many machine learning models. Bioinformatics. 2014;30(22):3152-3158. doi:10.1093/bioinformatics/btu520. PMID:25086004. PMCID:PMC4221122.

Documentation

Links