Asymmetric boosting

Abstract

A cost-sensitive extension of boosting, denoted as asymmetric boosting, is presented. Unlike previous proposals, the new algorithm is derived from sound decision-theoretic principles, which exploit the statistical interpretation of boosting to determine a principled extension of the boosting loss. Similarly to AdaBoost, the cost-sensitive extension minimizes this loss by gradient descent on the functional space of convex combinations of weak learners, and produces large margin detectors. It is shown that asymmetric boosting is fully compatible with AdaBoost, in the sense that it becomes the latter when errors are weighted equally. Experimental evidence is provided to demonstrate the claims of cost-sensitivity and large margin. The algorithm is also applied to the computer vision problem of face detection, where it is shown to outperform a number of previous heuristic proposals for cost-sensitive boosting (AdaCost, CSB0, CSB1, CSB2, asymmetricAdaBoost, AdaC1, AdaC2 and AdaC3).

Calibrating AdaBoost for Asymmetric Learning

“…As for the other variants, they are all methods that modify the training algorithm. CSB0 and CSB1 [17] do not use confidence rated predictions and based on the results of comparative studies [9,10,15], the two variants are typically dominated by CSB2. Asymmetric-Adaboost [18] was excluded from said studies as being similar to CSB2.…”

“…However it is often regarded as skew-insensitive [15,17], meaning it is unable to handle asymmetric tasks. There exist many skew-sensitive AdaBoost variants, including AdaCost [2,17], CSB0, CSB1, CSB2 [17], Asymmetric-Adaboost [18], RareBoost [6], AdaC1, AdaC2, AdaC3 [16], CS-AdaBoost [9,10]. However, most of them are heuristic and as a result they lack the theoretical guarantees of the original AdaBoost [7].…”

A survey on face detection in the wild: Past, present and future

“…In [112], the above work was extended by proposing to balance the skewness of labels presented to each weak classifiers, so that they are trained more equally. In [113] a more rigorous form of asymmetric boosting based on the statistical interpretation of boosting [55] with an extension of the boosting loss, was proposed. Namely, the exponential cost criterion in (3) is rewritten as:…”

“…The method in [113] minimized the above criterion following the AnyBoost framework in [114]. The method wes able to build a detector with very high detection rate [115], though the performance of the detector deteriorates very quickly when the required false positive rate is low.…”

High Detection-rate Cascades for Real-Time Object Detection

“…where I(y = 1) and I(y = −1) are indicator functions [11]. Hence, (23) is an asymmetric boosting loss function that can be minimized in a manner similar to AdaBoost, by gradient descent on the space of convex combinations of weak learners.…”