How are the loss functions of Multi-Class classification and Multi-label classification different ? Please give examples.
There are no difference on what you are asking for.
Sometimes, we could use either the term class or label to describe that the data has been previously annotated. This means that, if we have the data labeled to more than two classes, and you are going to perform classification on it, then this is simply named either multi-class classification or multi-label classification.
Mohamed Elhadad , appreciate your response, but it is not accurate. Precisely, your not accurate statement is:
"if we have the data labeled to more than two classes, and you are going to perform classification on it, then this is simply named either multi-class classification or multi-label classification"
This is because multi-class classification and multi-label classification are totally different. In multi-label problem, a data instance may be associated with multiple labels. This is as opposed to the traditional task of single-label classification (i.e., multi-class) where each instance is only associated with a single class label.
Dr. Samer Sarsam
Dr. Samer Sarsam, performing the classification process depends on what application domain the data are labeled to serve in. If the data has multiple of labels that serve in different application domains, it depend on what you want to classify. hence, the labels that are out of interest could be used normally as a features in the used feature vector for classification tasks.
For example, if we have news data, and the data are labeled with respect to its domain(sport, health economic,etc) and at the same time, their sentiment is reported/calculated, and these news were written by some of writers in addition to some other information in the news data. Here, if the application is related to opinion mining, then the sentiment column will be the target for the classification while the other fields will be just features in the feature vector. If the goal is to detect who is the writer of the news document, then the target will be the authors column. while if we want to detect the domain of these news document, the n the target will be the domain column.
The data by its nature is multi-labeled data, even before being labeled; you could consider the date as a label if you want to detect the flow of some events and so on.
Mohamed Elhadad
The concept of labeled data (with a single of many labels) is different from issues related to multi-class and multi-label classifications that we are discussing in this post. In your last reply, you are discussing data characteristics, while we are discussing aspects belong to two classification types.
To make things much clear, pleas have a look at the definition of multi-class ( https://en.wikipedia.org/wiki/Multiclass_classification ) and multi-label classification ( https://en.wikipedia.org/wiki/Multi-label_classification ).
Cheers,
Dr. Samer Sarsam
Common loss function for multi-class is categorical cross entropy (also named as softmax loss) because the ground-truth of each data correspond to a single class. For example image represent a cat OR a dog.
Multi label means that you could have more label (classes). For example image represent a cat AND dog. Common loss function is binary cross entropy loss (also named sigmoid loss). Keep in mind that binary cross entropy loss, for its nature, doesn't assume 0 value when target is equal to prediction, but it assumes the smallest possible value (near 0 of course).
There are also multi class multi label problems. Such as animal class with cat and dog labels ... , and car class with suv, coupé, cabrio etc...
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification