Integrated learning is through architecture and combining multiple learning to handle an idea of the learning task, currently divided into two categories: Boosting and Bagging.
For any of the integrated approach, we want to learn out of the base classifier with high accuracy and diversity, accuracy-based classifier can be integrated to ensure the accuracy of the results, while the diversity of requirements can make it ensemble classifier having good generalization performance. but usually, when the group classification accuracy are high differences between them will be smaller.
1. Boosting right way to learn each group classified by a serial manner, to adjust the training based on the training data errors in a classifier learning new base classifiers sample weight or probability distribution (error classification will be given larger probability or higher weights, such that when administered in training the classifier is a group higher sample of interest). Thus, a method based on integration to Boosting, a strong dependency between the base classifier, to be strings line learning. AdaBoost is representative Boosting algorithm.
2. Bagging Methods of the original training data set obtained by sampling a plurality of different subsets, then the study group learner data in each subset. Bagging strategy is such that between the base learners as independent as possible, which usually larger differences between the ways random forest each group learning to learn an Bagging is a representative algorithm, a random forest in the training data not just random sampling, and constructing a set of attributes is also randomly selected sub-tree set learning, introducing further randomness.
Variance and deviation from the point of view of learning Boosting method of decreasing the error in the training model, and Bagging method is more concerned with reducing the variance of the model.