Calculating the optimal number of bins for severely skewed data I have a data set with a sample size over three million numeric values. Close to 20% are either 0 or 1, with the maximum being nearly 18500. So the data is clearly quite heavily positively skewed.
I am trying to categorize some of this data by putting it into bins of equal width, so I decided to try and find the optimal number of bins. Using the Freedman-Diaconis rule it gave me a value of 126044.0262335108, this is clearly a ridiculously large number of bins for the data. 
Breaking the set into the Inter-decile range also proved fruitless giving me [0, 1, 1, 2, 3, 5, 8, 17, 47]
Reading elsewhere the square root of the sample size was suggested, this gave 1732.05081 which is more reasonable. However the method is quite crude.
I also looking into Doane's formula given here. But reading up on this method it seems to have been based on an incorrect hypothesis.
How should I deal with this level of skew in the data?
What is the best way to categorize this data? 
 A: Assuming that your goal is to visualise your data, no binning can allow you to appreciate the distribution in the range 0-47 and the remaining cases up to 18500. Even if you can fit the 0-47 range in a single cm of paper, the maximum (18500) will lie over 3 meters away. If you have a very large support, like a corridor wall, you can follow Nick Cox's comment and draw spikes at every value to convey a sense of the real scale of your data - in fact astronomers and science museums often do so to convey the real scale of the Solar System.
If your histogram needs to fit in a piece of paper or a screen, I would transform the variable as PtrZlnk suggests in his answer. Transformation log10(X+k) could work, and square or cubic root and asinh may work, too.
Once your variable is transformed, you could apply any standard binning strategy to the transformed variable, but don't forget to label your axes in the original variable.
A: If this is machine learning problem you can just try even 50 different binning strategies and cross-validate which one is the best.
If this is rather hypothesis testing problem think of it in cathegories of statistical certainty about value of predicted variable in bin and choose such a bin size that can quarantee that the value ascribed to bin is choosen with precision high enough. This however require more information about your modelling problem to define the exact algorythm of choosing right bin size.
Another idea is to trying to take logarithm od the variable, thus however would require adding some positive number to the variable.  Then it is more likely that relation to dependent variable would linear.
