Suppose we have a lot of machine learning algorithms (which can be any one we've learned before), can we use them at the same time to improve the performance of our algorithms? Also namely: Three Stooges equals.
There are several ways to aggregation:
Some of the poor performance of machine learning algorithm (weak algorithm), how to aggregation, become a better performance algorithm? Take a look:
We can see that sometimes aggregation performance is like doing feature transform, and sometimes it's like doing regularization.
Blending:uniform Blending, Linear Blending, any Blending
We can see: The performance of machine learning algorithm A is divided into two parts, performance of consensus (bias) and expected deviation to consensus (variance). and uniform blending improve performance by reducing variance, to obtain a more stable algorithm to achieve.
where α is bound to be greater than 0, the binding conditions can be removed.
Bagging
We can see that aggregation works because of the diversity of machine learning algorithms. So, how do you generate enough machine learning algorithms? There are several scenarios. Now we will focus on: diversity by data randomness.
We have previously imagined this situation in uniform blending. However, it is in the ideal state, 1) Our t can not infinity, 2) Our d is not infinite, now we use the following techniques to solve:
Random Forest
What is the random forest? is a special case of bagging: G is the case of a decision tree .
Why is it? Before we said uniform blending is to improve the algorithm performance by reducing the variance and making the algorithm stable. And bagging is a special form of blending. And we know that decision trees are sensitive to data, and different data can cause huge changes in the algorithm. Bagging just can reduce variance.
So it can be said that the random forest is a special case of bagging, it can also be said that the random forest is to improve the decision tree performance (stability) and a strategy used .
So what are the so-called "bootstrap" steps? Generate a lot of "D"?
How many decision trees does it take? The author used 12000 trees in one match.
Out-of-bagging (OOB) technology
Bagging technology we talked about this before:
That is, for a certain G, nearly one-third of the data is not used! It's a huge waste! How do I use these OOB data?
Recall validation:
Feature Selection
Assuming that each sample has a lot of feature, there are many redundancy features, there are many features that are not related to the problem, how do we choose the features we want?
Aggregation (1): Blending, Bagging, Random Forest