"finely" labeled classification problem I came across a kind of classification problem.
Suppose I have a dataset with labels, where each label is one of a1, a2, a3, b1, and b2. I want to make a classifier that estimates "coarse" label, a or b.
Here is a concrete example. Imagine a set of animal images labeled by the its species. We want make a classifier that predicts big classes, like mammal, reptile or fish. 
Q1. Does anybody know the name of this kind of classification problem setting?

To solve the problem, we can think of several approaches.


*

*First make the label coarse (a1->a, a2->a, b1->b,... ) and solve it as a standard classification problem.

*First solve the standard classification problem, predicting a1, a2,... Then make the label coarse.

*First solve the standard classification problem, predicting a1, a2,... (level-0 classifier) Then use the output scores of level-0 classifier as input, and make a level-1 classifier which outputs a, b,.. (stacked generalization approach)
Q2. What kind of approach is good for this kind of problem?
I guess the approach 1 is not optimal since part or label information is thrown away. As Marc Claesen has pointeded out, this approach 1 is good if there is an example that is obviously a member of a but not obviously a member of one of a1, a2,...
 A: 
I guess the approach 1 is not optimal since part or label information is thrown away.

Not necessarily. It depends on the data. Consider the following scenarios in which you would use opposite approaches.
If all subclasses per general class ($a_1$, $a_2$, $a_3$ for $a$) are very similar, it is likely easier to solve the problem of $a$ vs $b$ directly. Intuitively, there may be instances that clearly belong to $a$ but may not correspond as clearly to any of its subclasses.
In contrast, if the subclasses ($a_1$, $a_2$, $a_3$, ...) are dissimilar, it may not make sense to attempt to model the encompassing class $a$. 
A: You can use predictions (preferably real values, not binary) from model that uses all labels as input for prediction of final coarse labels.
Something similar to stacked generalization.
A: I believe this is called hierarchical classification and has been studied. One approach is to use multi-task learning - i.e. use a decision tree or neural net in order to try and take the hierarchical information into account. For instance if you use a neural network, you could have your outputs be the fine class labels as well as the coarse labels. Then, as in mult-task learning, the latent features that will develop in the hidden layer will benefit both tasks - these features may not develop otherwise. This will only work though if there is some "common information" shared between the fine labels of a coarser class.
