9
$\begingroup$

Suppose that I am interested in three classes $c_1$, $c_2$, $c_3$. But my dataset actually contains several more real classes $(c_j)_{j=4}^n$.

The obvious answer is to define a new class $\hat c_4$ that refers to all classes $c_j$, $j>3$ but I suspect this is not a good idea since the samples in $\hat c_4$ will be rare and not very similar to each other.

To visualize what I'm trying to say, suppose I have the following two variable space and the classes $c_1$, $c_2$, $c_3$, $\hat c_4= \bigcup_{j=4}^n c_j$ are depicted in red, til, green and black respectively. This is how I suspect my data would look like.

enter image description here

Is there any standard way to approach this problem? What would be the most efficient classifier and why?

$\endgroup$
4
$\begingroup$

I would use a two-step approach, using the idea of the $\hat{c_4}$ class you mentioned.

In the first step, use a binary classifier(trained on the whole dataset) to decide if a sample belongs to the class $\hat{c_4}$ (i.e. in any non-interesting class). For this, step you could also take a look in outlier detection methods, if the samples belonging in the "interesting" classes are much different than the rest.

If the result is negative, move on to the next step, a new classifier trained only on samples belonging in the classes $c_1,c_2,c_3$ and use that prediction as your final one.

I think that even using a simple clustering approach as a first step (e.g. 4-clustering k-means using as initial centroid values the average centroid $cent_j = \frac{\sum\limits_{x_i\in D: y_i=j}x_i}{\sum\limits_{x_i\in D: y_i=j}1}$ for each $c_1,c_2,c_3, \hat{c_4}$), would still be useful.

$\endgroup$

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.