Bayes regularization in the selection of weight coefficients in the predictor ensembles

Abstract

The supervised learning problem is discussed in the article: it is necessary to restore the dependence that maps a vector set into a scalar based on a finite set of examples of such a mapping - a training sample. This problem belongs to the class of inverse problems, and, like most inverse problems, is mathematically incorrect. This is expressed in the fact that if you construct the solution using the least squares method according to the points of the training sample, you may encounter retraining – a situation where the model describes the training set well, but gives a big error on the test one. We apply the approach when a solution is sought in the form of an ensemble of predictive models. Ensembles are built using the bagging method. Perceptrons and decision trees are considered as basic learning models. The final decision is obtained by weighted voting of predictors. Weights are selected by minimizing model errors in the training set. To avoid over-fitting in the selection of weights, Bayesian regularization of the solution is applied. In order to choose regularization parameters, it is proposed to use the method of orthogonalized basic functions, which allows obtaining their optimal values without using expensive iterative procedures.

Keywords

• обучение с учителем
• бэггинг
• некорректные задачи
• байесовская регуляризация обучения