Doubts about the algorithms in MEKA

Waikato / meka

Multi-label classifiers and evaluation procedures using the Weka machine learning framework.

http://waikato.github.io/meka/

GNU General Public License v3.0

200 stars 76 forks source link

Doubts about the algorithms in MEKA #42

Closed alexgcsa closed 7 years ago

alexgcsa commented 7 years ago

Hi,

I have three different (conceptual) doubts about the meta-algorithms. I am trying to understand them.

1) It is said in the documentation that Classification Maximization (CM) is a hard version of Expectation Maximization (EM). What does it mean this statement? What are the differences between them?

2) The Meta BR (MBR) has the description saying that "BR stacked with feature outputs into another BR". I did not understand it. What do you mean with this description?

3) In the description of Subset Mapper, it said that this algorithm "maps the output of a multi-label classifier to a known label combination using the hamming distance". Would it be the hamming distance between what?

Thanks in advance!

Alex de Sá

jmread commented 7 years ago

Hi,

CM does not take into account posterior probabilities (and thus may be used when none are available). This method is not specific to Meka, More details can be found in the literature.
In MBR the label predictions of a BR method become the features for a new BR.
In Subset Mapping, it refers to the Hamming distance between a prediction on a test instance and label combinations (i.e., label subsets) from the training set, it is 'mapped' to the closest one.

Cheers,

Jesse

alexgcsa commented 7 years ago

Thanks, Jesse.

I was also checking the code of Four-class pairWise classification algorithm (FW) and I didn't understand so much one of the methods (distributionForInstance) :

https://github.com/Waikato/meka/blob/master/src/main/java/meka/classifiers/multilabel/FW.java

I was wondering if the result (output) of this method should be the probabilities of the labels. It seems that is just the countings for each possible class and outputting the countings. I got a little confused at this point (I changed the name of issue because of it).

Thank you so much again.

Best regards,

Alex de Sa

jmread commented 7 years ago

Your understanding of the FW classifier is correct. Probably it would be nicer to divide the counts by a sum to obtain a kind of approximation of the probability (between 0 and 1) for each label. But it is not a requirement, since an appropriate threshold will take care of that.

Regards,

Jesse

alexgcsa commented 7 years ago

Alright. I understood what you said. I will possibly do that in the code then.

Best regards,

Alex de Sa