svm is for binary classification problem, if you want multi-classification, is nothing more than a few multi training svm chant
OVR (one versus rest)
For k classes (k> 2) case, the k svm training, wherein for determining the j-th svm arbitrary slice data belongs to the category j or non-class j.
Predicted time, has a maximum value of \ (w_i ^ Tx + bi \ ) indicates that the sample belongs to class i.
Suppose there are three categories of samples, A, B, C, need to be trained SVM 3, denoted as s1, s2, s3
Then outputs a sample x, go through s1, s2, s3, compared max (s1 (x), s2 (x), s3 (x)) of the category
OVO (one versus one)
For k categories, pairwise combinations, training k * (k-1) / 2 two svm, each sample is determined svm only for a specific k belonging to two categories.
Predict the time to do K * (k-1) / 2 times predicted by k * (k-1) / 2 Ge svm, with the vote determines how the sample belongs to that category.
Also assume that there are three categories of samples A, B, C, need to train 3 * (3-1) / 2 = 3 SVM, respectively SAB, SAC, SBC
Then enter a sample x, made prediction test 3, (AB, AC, BC), the results are assumed: B, A, B as the final category B
SVM Summary
Feature
- Focus on looking for the best dividing line (margin), to reduce the over-fitting (not sensitive to outliers, considering only supports vector)
- Kernel trick applications that can efficiently process SVM linearly inseparable scenes
Advantage
- Theory perfect
- Convex Optimization and dual (the KKT)
- Max Margin
- SVM objective function
- SVM dual form (lagrange)
- Slack SVM
- Kernel SVM
- SVM algorithm for solving the SMO
- Kernel support different functions
Disadvantaged
- Especially when the number of large, slow training
Overall, as jerry chiefs often talk about, this convex optimization (dual), kernel function such technology, not just for SVM, many places can also be a way. I feel SVM, which I suppose I really learned to the point, the hard-core technology.