However, I have read this amazing blog about binning data this whereby the author claims that adaptive binning is better than fixed-width binning. I understand the idea behind this because the some of the bins that we define in the fixed-width approach may have too little data distribution in comparison to other bins, which won't be a fair game to play whereby the adaptive approach, motivated by the idea of quantiles is better. Are there any more arguments, or more in depth analysis, of this hypothesis ?
I took a look at that link. It's pretty informative. As you already know, The cut function is used to specifically define the bin edges. There is no guarantee about the distribution of items in each bin. In fact, you can define bins in such a way that no items are included in a bin or nearly all items are in a single bin. The qcut function is slightly different. The qcut function tries to divide up the underlying data into equal sized bins. The function defines the bins using percentiles based on the distribution of the data, not the actual numeric edges of the bins. In conclusion, if you want equal distribution of the items in your bins, use qcut . If you want to define your own numeric bin ranges, then use cut.