Skip to main content
Figure 2 | BMC Bioinformatics

Figure 2

From: Improved variance estimation of classification performance via reduction of bias caused by small sample size

Figure 2

Unbiased estimates of m d1 (top) and σ d1 2(bottom). The true values of m d1 and σ d1 2 (dotted) are compared to the average of 50 conservative point estimates of m d1 and σ d1 2 (solid) for the 2-dimensional normal distributions used where N d = 100. Also displayed are two-sided 95% CIs, based on histograms of the 5000 estimates for different values of the test bag size N T1 . The true m d1 and σ d1 2 were obtained by testing 10,000 independently designed classifiers using 100,000 new, independently generated, test samples from the two normal distributions used. The estimates of m d1 are unbiased and the estimates of σ d1 2 are unbiased for N T1 ≥ 100.

Back to article page