We have, supposed, 100 classes. Each has 100 samples.
If we use two-class svm in training. The training must repeat 100 times to classify in one-vs-the-rest paradigm. If we take all samples of the rest in training, the performance is always the worst.
How to choose negative samples?
I think you must use ANN (Artificial Neural Network)
ANN can solve this problem earlier than SVM
even in ANN, if you put too many negative samples, you still meet the problem.
You can balance your classes assigning them weights inversely proportional to the size of the class. Popular LIBLINEAR package supports it with -wi option. For more complicated strategies e.g. sampling a subset from the negative class, see
http://ieeexplore.ieee.org/document/4695979/ or
http://www.sciencedirect.com/science/article/pii/S0167923609001754
I am not an expert, just searched these papers in Google Scholar.
Hi Hung,
You need to perform learning process under the evaluation of e.g., cross-validation (whole training set including positive and negative instances--no need to divide it yourself though).
HTH.
Samer
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training