Active learning is a family of strategies aiming at selecting unlabeled samples to label them and add them to the training set so as to improve the predictions. So it is supervised by construction.

Here is an example: suppose you want to train a predictor and the training procedure is quadratic in the number of samples (typically, kernel SVM). Instead of training on your full dataset which is very costly, your train on a very small subset which is much much faster. Then, you predict on the remaining samples and you use the active learning strategy to select a fraction of it and add them to the training subset, retrain and evaluate performances. Why doing that? Two reasons: First, it's very likely that performances are going to be the same when training on few well chosen samples and when training on the full dataset, while the former is much faster. It's even proven for SVM. Second, many training procedures have a warm start which means that adding a few samples to the training set and retrain the model is fast.

There is a second situation where you might want to to active learning, and that in the case where data are not labeled by default, but you have an oracle to label them on the fly. Typically, think of a interactive system with a human in the loop. You have the samples, do the predictions, present them to a human who labels some of them (relevance feedback loop) and use these labeled samples to update the predictor. In that case, you cannot present all predictions to the human, and thus you want to carefully present to ones which once added to the training set will enhance the predictor the most. This selection procedure is active learning, but it is still supervised, because the human has to label the samples before they are added to the training set.

Finally, if you training procedure converges to a local minimum of the objective function, you might want to use active learning as a way to improve the solution: instead of training directly on all samples, you select the ones that are more likely to lead to a satisfying solution. But this is more known under the name curriculum learning.

As you can see, all of this is independent of the prediction task and could be used for classification, regression, density estimation, ranking, you name it. The bottom line is, if you don't deal with interactive systems with relevance feedback loops or large enough datasets that you cannot train on them without a trick, then you probably don't need active learning.

If you are interested in using unlabeled samples without the oracle to label them, they you probably are referring to either transductive learning (prediction are valid only on the unlabeled data that were provided during training) or semi-supervised learning (you provide also unlabeled data, but the predictor works on the whole space). Of course you can do that for ranking, and still use active learning. However, semi-supervised (and transductive) learning is very tricky to achieve, very slow and do not expect miraculous results.

Dear David,

In my opinion, If the number of unlabeled instances are large. The model will bias for the method that calculates or guessing the heuristic labels and may also undergoes to high variance to unlabeled data and its calculation method. If the unlabeled instances are small, then learning algorithm can learn from labeled data and using the model as predictive model for unlabeled data. 

Thanks & Best wishes

Osman

Dear Samer

Unfortunately, Your feedback does not help at all

Best wishes, Osman