If I have a supervised learning system (for example for the MNIST dataset) I have features (pixel values of MNIST data) and labels (correct digit-value).

However sometimes people use the word target (instead of label).

Are target and label interchangeable? Is label just used for classification? Target both for classification and regression?


2 Answers 2


Target: final output you are trying to predict, also know as y. It can be categorical (sick vs non-sick) or continuous (price of a house).

Label: true outcome of the target. In supervised learning the target labels are known for the trainining dataset but not for the test.

Label is more common within classification problems than within regression ones. Nonetheless, they are often used interchangeably without great precision.

  • $\begingroup$ For clarification: If I would predict the price of a house (through this would be regression). Label would be the actual price of the house (the correct value I have for Training). Target the output of my ML system (both for training and inference). With the training I try to get my targets close to my labels. However my targets will probably never be perfectly correct. Did I understand correctly? $\endgroup$
    – Niklas
    Commented Jan 29, 2019 at 7:26
  • 1
    $\begingroup$ Just one thing in the last sentences: it is better to use "predictions" than "targets". You can think of the target as the main goal (house price), the label has having a known value (real price) and your predictions as the estimates (output of the model). $\endgroup$ Commented Jan 29, 2019 at 9:50
  • $\begingroup$ Just as an example where the target and label are different, there are training techniques such as label smoothing where you don't try to train the model to learn the exact label. For binary classification, the true label might be 1.0 and we might train the model to try to predict 1.0. In this case, the target is 1.0 as well. However, with something like label smoothing, the true label would be 1.0, but during training we would give a target of 0.9. This label smoothing is usually intended as a form of regularization. $\endgroup$
    – golmschenk
    Commented May 6 at 18:11

I will give an example where they are not interchangeable.

Label and target both can express the meaning of y depending on x, but only label has a meaning of describing the input, for example:

In image classification: a training example (cat image pixels, cat), we can say the cat is the label of this image because it's just describe the kind of this image.

But in word2vec in which we use current word to predict its context word: a training example, say (orange, juice), here we may not say juice is the label of orange, very strange right? so in this scene, we can only say juice is the target of the orange, only express it's the dependency.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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