3
$\begingroup$

Background: I am working on the problem of classifying objects found in some biological images. Time and again, we encounter objects which do not fall into any of the categories/classes we are interested in and we would like to filter out these uninteresting objects before analyzing the interesting ones. I look at this as an outlier detection problem. I have prepared a training set with inliers (all interesting object categories pooled together) and outliers (objects we are not interested in). Visually, the outliers have a different (but very diverse/surely-multimodal) appearance characteristics when compared to the objects in the inlier class (which is also intuitively multimodal). So, I have computed some features that quantify an object's appearance in various aspects including texture.

Now, I am faced with the question of whether to solve this as a two-class classification problem or a one-class classification problem. My training set is quite imbalanced, with 3500 examples (~96%) from the inlier class and 150 examples (~4%) from the outlier class.

Question: I want to ask at which degree of class-imbalance between the positive and negative classes would you strongly prefer solving such problems as a one-class problem as opposed to a two-class classification problem?

$\endgroup$
1
$\begingroup$

It is hard to specify such a degree exactly. But in general when you have a very unbalanced class distribution (as in your case) it is better to treat the problem as an outlier detection problem. This is because since there is too little information about one class the classifier cannot learn the true boundary between negative and positive instances.

However, you can always do cross-validation experiments to see whether treating the problem as an outlier detection problem or as a two-class classification problem works best.

$\endgroup$
  • $\begingroup$ I would, also, suggest considering pre-processing the data distribution before applying any classification model. You may apply random undersampling or oversampling or synthetic approaches such as SMOTE. One-class classification is still an option that can challenge previously suggested solutions. $\endgroup$ – soufanom Aug 26 '13 at 2:58

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.