Basically, I am doing my project in which I need to classify the rice plants as healthy or they lie in 9 of the diseases. So I have used multi-class SVM for the classification.
The steps for my project included:
pre-processing ---> Segmentation ---> Feature Extraction (I extracted a total of 13 features based on the texture and color) ----->then I have prepared a training matrix based on the features and then I trained the model. So I wanna ask how can I find the accuracy of My SVM Model?
https://www.mathworks.com/help/stats/classificationecoc.html;jsessionid=7b5758ad4aeb601024cfbd4d35f8
Dear Sadia, The detailed analysis is given in the link listed below:
Multiclass model for support vector machines (SVMs) and other classifiers
https://www.mathworks.com/help/stats/classificationecoc.html;jsessionid=3e6a0565813c7e2474560c8f75d0
It depends on how you define accuracy. I tend to think of accuracy as how well the probability prediction matches the frequency of observations in reality. E.g., does a prediction of 50% class A result in the truth class being A 50% of the time this prediction occurs.
Many metrics exist to capture this relationship of prediction vs observation. Brier Score decomposes into its component measures (one of which is accuracy). KLD decomposition into its equivalent component measures. The attached KLD paper uses the term "reliability" but its accuracy--this is a single number to gauge an overall performance against all classes. You can also plot prediction vs observation too for each class on one plot so you can visualize the accuracy. For the visualization you have to keep in mind the prior used, which sounds like could be 1/10 for all classes if the prior is uniform.
Another common definition for accuracy is how often the predicted class call matches reality--regardless of the associated probability. A confusion matrix is best for this kind of evaluation, which gives you a class-wise accuracy measure for class calls. I like normalizing on prediction instead of truth to get the power (in the truth table sense) to represent this kind of accuracy.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification