# Using probabilities for multi label classification

After working for a while on this text classification problem, I realize that some documents belong to more than one class. I am using multinomial logistic regression which also provides a probability distribution over the classes (labels). I wonder if it is a good idea to use this distribution for multi labeling. For example, when the probabilities are [0.3, 0.6, 0.1] for the classes A, B, C respectively, I can label the document with the classes that have a probability for that document higher than a predefined threshold (say 0.25) .

Is this a good idea? I've made a Google search but couldn't found any document mentioning a method similar to this. How reliable is this method? What do you think?

To be more clear about my problem space, there are like 20 classes and mostly a document belongs to either one or two of these classes.

Multinomial assumes that an outcome belongs to only one class, but you can redefine classes. E.g. if there are two original classes A and B, then you can label the documents as belonging to three mutually exclusive classes:

I - document is A only

II - document is B only

III - document is both A and B.

• Yes, that is one of the approaches. But I have nearly 20 classes, which could lead to many combinations. – hrzafer Nov 10 '16 at 19:22
• If you have no use for that many classes, then collapse a few classes into one. E.g. I, II and III can be collapsed into a single class, "A or B". – Nik Tuzov Nov 14 '16 at 20:31