It is easy to confuse the two probabilities of multi-classification problem and multi-label learning, especially when the two concepts are put together, and when it comes to the use of many-to-many models for multi-classification problems, it is more likely to be dizzy, so the author is simple here Introduce the similarities and differences between the two.
1 Multi-classification problem
1.1 One to One
When the classification is multi-classification (that is, k>2), the multi-classification problem can be flexibly handled as a normal 2-classification problem , which is a pairwise combination of different classes . For example, when k=3:
The total number of classifiers is: k(k-1)/2=3, which means that 2 of the k classifiers are always selected.
The classifier situation is ( positive class|negative class ): (class 1|class 2), (class 1|class 3), (class 2|class 3)
1.2 One pair of others
A pair of other models looks at other categories as negative classes, which can reduce the number of models. The number of one-to-one models is n(n-1)/2, and the number of other models is n.
The classifier situation is ( positive class|other classes are negative class ): such as k=3, (class 1|class 2, class 3), (class 2|, class 1, class 3), (class 3|class 1, class 2)
1.3 Many to many
The many-to-many model is a further improvement on the other models, because a pair of other models may encounter a situation where the output of 2 or more classifiers is positive. At this time, a pair of other models selects the classification label with the highest probability for the sample , And many-to-many modelers give both classification labels to the sample. This overlaps with multi-label learning. The first two situations are better understood. Under normal circumstances, we discuss the first two model situations by default, that is, one pair One, one other. Just remember this special case.
2 Multi-label learning
Multi-label learning is also called multi-label learning. It is to assign multiple tag information to a picture. As in the picture below, its label can be (tree, sky, person),
3 Similarities
The multi-classification problem is similar to the multi-label learning in that it is to give some kind of label information to the sample, especially when the multi-classification problem is a many-to-many model, the multi-classification can include multiple labels.
4 Differences
Generally speaking, a multi-class problem is to select the classification label with the highest probability for the sample. When doing multiple-choice questions in the previous exam, choose the most correct one to give it. It is a single-choice question, while multi-label learning is a multiple-choice question. The label information suitable for the sample is to select the suitable label. It is a multiple choice question.
From the perspective of machine learning, multi-classification is a concrete problem, and multi-label learning is a learning framework, a general paradigm, and an abstraction of concrete.