Over the years, my go-to techniques when dealing with imbalanced classes have been:
-
oversampling the under-represented class: I used to think that this method would always yield overfitting but that's not the case. This paper that I discovered thanks to fastai courses changed my mind with respect to oversampling. The paper compares different methods of dealing with imbalanced data in the context of CNN. Here are the conclusions of the paper
-
Tweaking loss functions or using loss functions that are made for this purpose like focal loss (Yes, I use it even on tabular data and not only in CV)
-
Generate new data using GANs but this often ended-up with a failure, but who the hell doesn't like playing with GANs!
I have to say that in most day-to-day tasks, I prefer using the first two methods as they are easy to implement and not energy-consuming^_^
To make sure there aren't some handy-dandy techniques that I've missed, I've decided to go through Soledad Galli course on Imbalanced data.
I've been writing down notes while taking Soledad Galli course. The course is easy to follow along. What I didn't like about the course is that it feels like someone's throwing a bunch of methods at you but no thorough comparison is made between the methods. If you haven't tackled the issue of imbalanced data before, this course might be really good for you. Otherwise, it might be a bit boring as was the case for me (but it has more to do with the fact that I knew many of the methods in the course. Nonetheless, the author applies different methods to a wide range of datasets.
The notes are not a real summary of the course, if you don't know the concepts discussed in the course, you'll have a hard time understanding the different methods by simply going through the notes.
You can find my notes here: Notes
- Machine Learning from Imbalanced Data Sets 101: This short paper tries to give a definition for the problem of imbalanced datasets when it comes to ML models. To do so, the author asks a number of questions and discusses in a very high-level the methods used today. The paper dates back to the year 2000. Yet, the discussion and the arguments remain valid.
As I've read many blog posts, I've decided to list only the one that I did really like
Here is the list of the datasets used in this repo:
- kdd2004.csv: This file's data comes bundled with imbalance-learn package. The original version was pre-processed before being saved to the csv file.