A classifier can be correct without winning decisively. In boosting, a margin measures the combined classifier’s support for the correct class relative to a competing class. It adds information about the strength of the separation that training accuracy alone does not capture. section 8.8.3
A boosted classifier combines many individually limited models. AdaBoost can continue increasing margins even after it reaches perfect training accuracy. This offers one explanation for continued improvement on new data despite no further reduction in training error; it is not a guarantee of generalization. section 8.8.3
A related vote margin
In a random forest of 100 trees, suppose the known class is A and the votes are 82 for A, 12 for B, and 6 for C. The margin is 0.82 − 0.12 = 0.70: the vote proportion for the known class minus that of its strongest incorrect competitor. A 51–49 vote also predicts A correctly, but its margin is only 0.02. This is a forest margin, not a geometric distance or a specified boosting calculation. section 9.4.1 section 8.8.3
The distinction matters when comparing models. An SVM’s margin is geometric and enters its training objective; a forest margin describes the voting advantage for a known class. Boosting margins likewise describe combined support rather than the SVM’s distance from a boundary. These notions share an interest in separation beyond mere correctness, but they are not interchangeable. section 8.8.3 section 9.4.1 section 9.7


