Can someone explain to me, why association rule learning is considered a supervised learning approach? The way i understood it is that the algorithm takes a bunch of coherent data-sets and computes associations based on these sets:
{a, b, c}
{a, b, d}
=> a -> b
=> b -> a
The way I see it, there are only arbitrary data sets. No specific target vectors. Why is this called supervised?
I guess it is an open discussion if one considers Association rule learning as an unsupervised or an supervised learning task. While Wikipedia counts it to the group of supervised learning algorithms other resources count them to the class of unsupervised learning algorithms:
As opposed to decision tree and rule set induction, which result in classification models, association rule learning is an unsupervised learning method, with no class labels assigned to the examples.
Machine Learning and Data Mining - Springer
I suppose it comes down to how the actual learning part is implemented. One could create a dataset of training data - label pairs such as in your example:
{a, b, c}
{a, b, d}
=> a -> b
=> b -> a
Having a couple hundreds or thousand of these pairs one could train a Neural Network to understand the underlying patterns in the dataset with fairly good accuracy as I would suppose. This would then be a Supervised Learning task, where the NN learns from pre-calssified examples.
If on the other hand the algorithm is implemented in such a way that the associations are computed based on: Support - Confidence - Lift - Conviction it would be an Unsupervised Learning task.