I have 4 groups of particles with different features. For instance particles in each group have different colors: red, white, black and blue.
I trained (with four single classes) a supervised machine learning algorithm (e.g. SVM) for classification.
For testing, when the particles contain single feature (only one color), the classifier returns its label correctly. But, when the particles are mixed (e.g. red & blue), the classifier returns either it is red or blue. However, I want that the classifier recognize this group as red & blue (both of them).
Note: the mixed red&blue features is not considered as a separate class to train the classifier.
I really appreciate your help in advance.
I think the question is a bit unclear. Are you using your features as classes as well (i.e. one feature is color=red, then class=red)?
If that is not the case: this is a multi-label classification problem (one example could have more than one label assigned, or can belong to more than one class). For this you can train individual SVMs using a OnevsAll approach, where the positive examples (particles) are the ones that identifiy a specific class (e.g red), and the negative examples are all the examples in the rest of the classes (e.g. white, black, blue). With this you will have 4 individual classifiers that can determine if a new particle belongs to a specific class or not. Each new test particle goes then to each of the trained classifers and they will decide if the particle belong to each particular class and assign the corresponding labels.
I did a review of several classification details in the following paper.
Chapter A Survey of Automated Hierarchical Classification of Patents
Thanks for your answer Juan Carlos Gomez. I train the classifier using four groups (classes) of features where each group represent different colors of particles.
Thanks for your explanations on OnevsAll approach. I usually use OnevsOne approach and have not tried OnevsAll which might be the solution as you mentioned.
I can imagine neural network classifier with 4 output sigmoidal neurons (one neuron for each color). The target output vectors are: [1 0 0 0] for red, [0 1 0 0] for white, [0 0 1 0] for black and [0 0 0 1] for blue. You train the network using your training set. After training you present input sample and observe net output. If you get output vector [near 1, near 0, near 0, near 1] you can infere that the sample is similar to both class red and blue.
Or different approach. You build 4 classifiers distinguishing one class from others. You can use SVM. So 1st classifier returns 1 as output for class red, and 0 for other classes. 2nd classifier returns 1 as output for class white, and 0 for other classes, etc. After learning you present input sample to all classifiers. When 1st and 4th classifiers return 1, and others return 0, you can infere that your sample belongs to both red and blue class.
Let me know about the results.
Good luck.
You can try hamming neural network. And set the target {blue, red, blue&red} .
For me it seams like the kind of problem that you can easily model using fuzzy logic. Have you tried fuzzy classifiers or neuro-fuzzy approaches?
Thanks everyone for your answers. They are really helpful.
No Nathalie I haven't tried neuro-fuzzy approaches before.
I am trying to apply different algorithms that can tackle this issue. Except LibSVM, is there any reliable package or toolbox available in MATLAB?
- Badges
- Science topic