7
$\begingroup$

Until recently, I thought that "labeling" and "classification" are synonyms. But when I started another question about terminology in computer vision I thought about it: Is there a difference between "labeling" and "classification"?

I thought that the "class" is the concept you want to detect and "label" is what you assign to data. So the "class" is a concept which leads to the data and "label" is only the name. Hence "labeling" would be the same as "classification" as both want to make a statement about the underlying class which lead to the data.

Articles

A quick search via Google Scholar revealed that some articles use both terms in the title:

  • Markus Eich, Malgorzata Dabrowska, and Frank Kirchner: "Semantic Labeling: Classification of 3D Entities Based on Spatial Feature Descriptors"
  • Chunlin Li, Dmitry B. Goldgof, and Lawrence 0. Hall: "Knowledge-based classification and tissue labeling of MR images of human brain"
  • Ray Blanchard: "The classification and labeling of nonhomosexual gender dysphorias" - another research area but probably it is the same difference between the two words?

So I guess there is a difference between "labeling" and "classification". What is the difference?

Google N-Gram

enter image description here

classification seems to be a much boarder term.

$\endgroup$
2
$\begingroup$

Totally disagree with @Derek Janni. Be careful about notation however you should not get lost in terminology. Those papers you mentioned used the term "labeling" literally but in Machine Learning/Data Mining community, labeling is the process of preparing data for supervised learning (classification)! It has nothing to do with the ML task!

Those papers used the term to show that after supervised learning they can recognize different labels of different objects so they used the term labeling but you probably can not find in any literature that they use these two terms as synonyms.

| improve this answer | |
$\endgroup$
1
$\begingroup$

The way I view it: 'Classification' (in the context of machine learning) is a type of problem in which you assign a 'label' to an object. Formally, 'Classification' is a type of problem whereas labeling is a function from an object to a set of labels (maybe infinite).

Much the same way Regression is a type of problem where you, again, assign a label to an object only this time the label is some real number.

Both in Classification and in Regression you are attempting to find the 'best' labeling function with respect to some metric/loss function.

| improve this answer | |
$\endgroup$
1
$\begingroup$

After having read a lot more papers and having talked to many people about machine learning topics, this is how I would define the words:

A class as an abstract concept which exists. Each class has properties and can have a lot of different labels. For example, the class cat has the properties "feet" (with the value 4), the property "Genus" with the value "Felis". There are many way members of the class can look like. Also many labels: cat, Katze, Felis silvestris, 🐱, 🐈.

A label is just a sticker you put on the concept. A name. We need a word to be able to talk about the concept.

I use labeling for the manual process of defining which parts of the dataset belong to which class. And I use classification for the process of the automatic classifier deciding which part of the data belongs to which class. So typically, labeling is done by a human and proceeds classification which is done by the machine.

| improve this answer | |
$\endgroup$
0
$\begingroup$

Short answer:

No, there is no difference between labelling and classification.

Class - a set or category of things having some property or attribute in common and differentiated from others by kind, type or quality. See 'category'.

Label - word or phrase indicating that what follows belongs in a particular category or class.

To classify something is to label it, they are the necessarily the same thing. The term labelling probably evolved because "label" allows you to avoid saying "class" which has other connotations in Computer Science.

Label is much simpler, and in all cases, classification is just the act of putting labels on objects (or learning to correctly do so).

The discrepancy you're seeing in the use of labelling/classification comes from the simple fact that a title like:

"Semantic Classification: Classification of 3D Entities Based on Spatial Feature Descriptors" or "Knowledge-based classification and tissue classification of MR images of human brain"

Sound really awkward.

Like most academic paper titles, these are just overly complex descriptions of what is in the paper that explain exactly what is going on without sounding redundant.

TL;DR - Don't get hung up on terminology!

| improve this answer | |
$\endgroup$
  • $\begingroup$ Your argument that it would sound awkward is strange. Of course, I would rather write "Semantic Classification of 3D Entities Based on Spatial Feature Descriptors" instead of "Semantic Classification: Classification of 3D Entities Based on Spatial Feature Descriptors". Writing a paper is usually a lot of work. I'm certain that people spend quite a bit of work in the title. But I guess I should simply try to get an author of one of those papers to answer my question. $\endgroup$ – Martin Thoma Nov 28 '15 at 21:06
  • $\begingroup$ My point was that it's generally regarded as awkward to repeat the same word multiple times in a title/sentence - hence the use of the word "labeling" instead of repeating "classification". Really a minor point, I guess. Perhaps the authors used "Labeling" as a search-engine trip to show up for queries on that topic, despite the fact that classification is the same thing. I personally like the title they decided to go with the best :) Also the proof is in the pudding, I'm assuming you've read the papers and noticed that what they are doing is classification at its core. $\endgroup$ – Derek Janni Nov 28 '15 at 23:11
0
$\begingroup$

Labels come up in conjunction with classification when the object does not belong to a single class but to a broader set, hence the term "multilabel learning" or "multilabel classification". Since they refer to discrete classes, they can be used synonymously, but I would recommend using the traditional terminology (classification when a single class is to be assigned) to avoid confusion.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.