In the first section of SVM we have a simple concept of generalization error. In SVM, the maximal solution of margin is to minimize the generalization error. So what is the generalization error, and why is the margin greatest when the generalization error is minimal? Now let's explore.
What is generalization error? The wiki gives such a description:generalization error (also known as the out-of-sample error) is a measure of how AC Curately an algorithm are able to predict outcome values for previously unseen data. This means that the generalization error is used to describe a model (algorithm) trained on the training set to predict the accuracy of the data outside the training set. Here is a simple example of the assumption that the school is going to choose one of the highest level students to participate in the National Physics Competition (the problem), then we need a process to select (training). At this time, the school teacher out of a 20 physics questions (training set) for everyone to do, the highest-scoring students (Model) selected to participate in the National Physics Competition (decision Rule). But is this one of the highest scores to represent the highest level of the entire school? That must not be, maybe he happened to do one of the 19 questions, but the actual level is not so high, the other physics problems will not do, if this is the case, we say that the generalization error is relatively large. If this assessment really chooses the highest level of students, then we say that the generalization error is relatively small. The generalization error is probably one such thing.
Support Vector Machine (3): Talk about generalization error (generalization error)