I am trying to understand how NN can be used for multilabel classification and I read the following paper: "Large-scale Multi-label Text Classification —
Revisiting Neural Networks" (http://arxiv.org/pdf/1312.5419v3.pdf) . Would someone please help me understand the following?
I don't understand how the training examples are transformed for NN. Did they use one hot vector with 1 at the place of relevant label.
Dear Aaqib Saeed,
Look the links. May be useful.
https://arxiv.org/pdf/1502.05988.pdf
https://users.ics.aalto.fi/jesse/talks/Multilabel-Part01.pdf
Regards, Shafagat
The MULAN tool offers a suite of algorithms for multi-label classification, including NNs. They also provide a collection of benchmark datasets: http://mulan.sourceforge.net/
It is also possible to use a problem transformation approach, such as BR or LP, along with NNs. Check these papers:
lpis.csd.auth.gr/publications/tsoumakas09-dmkdh.pdf
https://www.researchgate.net/publication/262323962_A_Genetic_Algorithm_for_Optimizing_the_Label_Ordering_in_Multi-label_Classifier_Chains
Conference Paper A Genetic Algorithm for Optimizing the Label Ordering in Mul...
Thank you Eduardo. I wanted to implement such a Multilabel classifier myself. I got the answer in the following paper: http://cse.seu.edu.cn/people/zhangml/files/TKDE'13.pdf
If I understand correctly I think the output probabilities (k-hot vector) from NN model (with binary cross entropy cost function) can be consider as multilabel output if certain probabilities are above some threshold.
Please correct me if I am wrong.
Yes, there are some methods based on NN, Naive Bayes and other techniques which employ a threshold-bases strategy
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Artificial Intelligence