Special Invited Paper-Additive logistic regression: A statistical view of boosting

By Jerome H. Friedman. 2000

6,512 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Statistics and Probability, Optimization & Training, Statistical Learning Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Statistics and Probability

In Optimization & Training

In Statistical Learning Theory

Citations in the corpus

Explaining Predictions from Tree-based Boosting Ensembles (2019, methodology): “(2) We extend our new method for adaptive boosting to gradient boosting [4], where we not only train in sequence but also train on the negative gradients of the previous tree.”
Reading Text in the Wild with Convolutional Neural Networks (2014, methodology): “The slidingwindowclassifier is an ensemble ofweak decision trees, trained using AdaBoost (Friedman et al. 2000), using the aggregate channel features.”
Geometric context from a single image (2005, methodology): “Both the label and the homogeneity likelihood functions are estimated using the logistic regression version of Adaboost [3] with weak learners based on eight-node decision trees [8].”
On integral probability metrics, φ-divergences and binary classification (2009, neutral): “(α)u−l(−α)). (12) Moreover, the relation in Eq. (11) holds whether or not l is convex. 7Choosing l(α) = max(0,1 − ) results in support vector machine [25], ) = log(1 + e−α) yields logistic regression [10], l(α) = e−α yields Adaboost [9] etc. 5 (b) (φ-divergence to l-risk) Given a lower semicontinuou”
Sharing Visual Features for Multiclass and Multiview Object Detection (2007, methodology): “We start with a brief review of boosting for binary classification problems [29], [28], [14].”
Empirical margin distributions and bounding the generalization error of combined classifiers (2002, methodology): “This becomes especially important because of the growing number of boosting type methods (see Friedman, Hastie, Tibshirani (2000), Fried-man (1999)) for which a comprehensive theory is yet to be developed.”
Node harvest: simple and interpretable regression and classication (2009, neutral): “Use of the L 2 -loss as a convex surrogate for misclassiﬁcation error is thus not only appropriate for NH, it is even beneﬁcial since it allows for an interpretation of the predictions ˆ Y ( x ) as weighted empirical node means.”
Recovering intrinsic images from a single image (2005, methodology): “To obtain Pr½D s jG, the probability of a derivative being caused by shading from the gray-scale classifier, we used the method suggested by Friedman et al. to transform the output of an AdaBoost classifier into the probability of each label [8].”
Recursive Aggregation of Estimators by the Mirror Descent Algorithm with Averaging (2005, methodology): “the classication theory, proving the optimality of the method of empirical classication error (1) (see [4,12,27]), and practice, where quite other approaches were applied, such as, for instance, SVM and boosting based on numerical minimization of convex functionals of the empirical risk dierent from”
Scoring diverse cellular morphologies in image-based screens with iterative feedback and machine learning (2009, methodology): “Overfitting is a concern when using machine learning algorithms, but boosting variants are fairly resistant to it (28).”
McRank: Learning to Rank Using Multiple Classification and Gradient Boosting (2007, methodology): “For example, it is well-known that, although one can use regression for classification, it is often better to use logistic regression especially for multiple classification [8].”
Vision‐Based Automated Crack Detection for Bridge Inspection (2015, background): “The details of the gentle boost algorithm can be found in the following references (Friedman et al., 2000; Torralba et al., 2004).”
Ensemble Tracking (2005, background): “This can probably be greatly accelerated if we sample the pixels (because the feature vector associated with each pixel already captures some local information) or if we ignore samples with low weight, as was suggested by Friedman et al. [12].”