Give you these fruit pictures and tell you which apples are. So now, let's summarize what is Apple?
1) The apples are round. We found that some apples are not round, some are round but not apples,
2) Where these violations of the "Apple are round" the rule of the picture, we get "apples are round, may be red or green". We find that there are still some pictures that violate this rule;
3) to the violation of the rules of the picture, we found that "the apples are round, may be red or green, and have a stalk."
This classification is complete.
Simulating this process is the adaboost algorithm.
Start with a weak classifier, then focus on the sample where the error occurred, get the next classifier. AdaBoost is characterized by:
1) different classifiers are serial trained, i.e.: The second classifier relies on the first classifier;
2) obtain a new classifier by focusing on the data that has been wrongly divided by the existing classifier.
How to focus on the data of the wrong points? By increasing the weights of the split data. This requires:
In order to have more physical significance, we would have to scaling the ε:
What is the physical meaning of this? When the data is wrong, ◇ must be greater than 1. Equivalent to the data weight amplification, while the data classification is correct, ◇ less than 1, the equivalent of reducing the data weight.
Another question: How to find weak classifiers? We know that any machine learning algorithm is possible, but here is a common weak classifier: a single-layer decision tree.
Aggregation (2): Adaptive Boosting (AdaBoost)