Multi-class classification and multi-label classification

Given a set of training instances (X1, Y1), (X2, Y2), ... (Xn, Yn), typically, each instance of Xi i=1,2,..., N is an m-dimensional vector, Yi is a vector with an L (l>=1) category, and the task of classifying is to learn a model f:x->y from the training instance, thus giving a trustworthy category prediction to the new instance.

The classifier for multi-class classification (Multiclass classification) is designed to specify a unique classification category for a new instance, with two common strategies: based on the posterior probability or distance given to all categories of measures at once, select the category with the largest measure as the forecast category , decompose the multi-class classification into many two-tuple classification questions, and then combine the results of all the two-tuple classifications.

The Multi-label classification (Multilabel classification) classifier assigns multiple categories to a new instance. This classification model has a wide range of practical applications, such as: A document may belong to multiple classifications at the same time, and a protein may have multiple functions. Also, there may be a certain dependency or binding relationship between multiple tags, such as Go (Gene Ontology), which is composed of all the functions of the protein. This dependency or constraint relationship has hierarchical characteristics, which can often be described as a tree or a direction-free graph structure, which the machine learning Community calls hierarchical multi-label classification. Because the output of the model has a hierarchical structure, the hierarchical multi-label classification belongs to another recently very active research area: structural prediction. Hierarchical multi-label classification and structural prediction are new and challenging research areas.