SVM 10 Q Ten Answer

What is a linear classifier?

The linear classifier attempts to derive a categorical hyper-plane from the sample in the training set, with the goal of maximizing the number of samples from different classes in the training set, and finally applying the categorical hyper-plane to the classification of the new sample.

What is the difference between SVM compared to a general linear classifier?

SVM to maximize the interval between different categories to optimize the target.

What is the difference between SVM and logistic regression?

• Logistic regression reduces the sigmoid information of those points away from the categorical hyper-plane through functions, and SVM ignores the information of those points directly.
• The logical regression output sample is attributed to a certain class of probabilities, and SVM cannot directly output this probability.

What are the benefits of maximizing the interval?

The interval is related to the structural risk of the classification problem, and maximizing the interval equals minimizing the structural risk, thus obtaining a better classifier.

What is structural risk?

Structural risk is the and of experience risk and confidence risk.
Usually when people build a classifier, the first thing is to focus on the classifier's classification error on the current training set, the smaller the error, the less the experience risk, the better the model to fit the experience. But the pursuit of the classifier in the training set of the best performance, it will bring the problem of fitting, resulting in the classifier in the training set unparalleled, encounter real problems but a mess of the phenomenon. Then people put forward a second indicator, used to measure the generalization ability of a classifier, called the confidence risk, plain English is how much I can trust this model, to help me solve the problem of new samples. However, the confidence risk is difficult to measure accurately, at present can only give an interval, and the simpler the model, the less the confidence risk.

What is a support vector?

The maximized interval forms a boundary on the set of positive and negative samples, and the sample on the two boundaries is the support vector. The classification of subsequent new samples relies only on the information of these support vectors, thus reducing the complexity of storage and computation.

Why change the form of the problem?

In order to make up a two-time planning problem, we can ensure the global unique optimal solution is obtained.

Why do you want to convert the Cheng Lagrange duality problem?

The conversion to duality problem can get more efficient solution, and also facilitates the introduction of kernel function.

How does SVM solve the problem of linear non-division?

The projection of the feature space to the higher dimension can transform the curve in the original feature space into a straight line, that is, we can obtain a linear problem in the higher dimensional space, and also judge the new sample in this higher dimensional space.

What is the function of nuclear functions?

After the feature space is ascending, the dimension is non-linear expansive (2-dimensional to 5-dimensional, 3-dimensional to 19-dimensional ...). ), it is easy to cause dimensional disasters. The kernel function allows you to calculate the decision boundaries of a high-dimensional space in the space of the original attribute.

How to deal with the situation of labeling errors?

Through soft edges (Soft margin). A relaxation variable is introduced for all constraints that do not satisfy the condition, and a penalty item is added to the target function. The multiplier factor C of penalty is determined by trial and error method in training.

How to classify?

Two ways: One is "the other", that is, each category and all the remaining categories as a positive and negative sample input, and the second is "one to the other", the 22 category after the comparison of what kind of, then take this category compared with the third category.

Original: http://guoze.me/

Thank you for your attention websites!

SVM 10 Q Ten Answer