Whenever I want to convert a continuous feature into categorical with bins, I use one of the following two ways.
- Freedman–Diaconis rule (wikipedia source)
Bin width = 2 x IQR(x) x n^(-1/3)
Where IQR(x) is max-min
and n is the number of your observations in your sample.
- Frequency Bins
You get such a bin width that each of them will have the same number of observations.
I don't have a rule of thumb, where to use each of them. I usually try both and measure the impact on accuracy. Then decide which I will use in the final model.