Instead of binary classification, I would like to classify the samples into class A and class B. if class A, then further into A1, A2, A3. Similarly B into B1 & B2. What classifier to use? Which tool?
Sounds like a decision tree. If you know the "structure" of your classes you could build the decision tree by hand, where at each node there would be an arbitrary classifier.
You could also have a look at hierarchical SVMs (http://www.csr.utexas.edu/hyperspectral/oldwebsite/yangchi/HSVM.htm).
http://www.csr.utexas.edu/hyperspectral/oldwebsite/yangchi/HSVM.htm
Some years back I developed an algorithm that autonomously builds a hierarchical classifier. This might not be what you are looking for since in my case you would specify only the lower level classes and the algorithm would autonomously construct the "aggregated" or higher level classes. You can check it out here: https://www.researchgate.net/publication/2474715_ARTD_Autonomous_Recursive_Task_Decomposition_for_Many-Class_Learning
Article ARTD: Autonomous Recursive Task Decomposition for Many-Class Learning
We developed a algorithm for semi-supervised clustering to look for clusters (micro-distributions) within classes. Here is the link.
http://ieeexplore.ieee.org/xpl/articleDetails.jsp?arnumber=6382192
See if it helps.
Hierarchical classification like you proposed is an open problem in machine learning. There exists some proposes like those bellow applied to biological data but the classifiers are developed for each specific domain:
The state-of-art propose:
http://www.researchgate.net/profile/Leander_Schietgat/publication/226246349_Decision_trees_for_hierarchical_multi-label_classification/links/0f31753bfe66a65c4a000000.pdf
A good survey about the subject:
http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.183.302&rep=rep1&type=pdf
Article Decision trees for hierarchical multi-label classification
Multiclass Support Vector Machines can be considered in same extent as a hierarchical model, for instance you can see the following paper available at researchgate.
We are currently developing the R library for its use.
Article: SVMTOCP: A Binary Tree Base SVM Approach through Optimal Multi-class Binarization
Diego Arab Cohen, Elmer Andrés Fernández
[Show abstract]
Lecture Notes in Computer Science 01/2012; 7441(7441):472-478. · 0.51 Impact Factor
- Badges
- Science topic
- Similar topics
- Filing