I am working with a pure categorical data set. And some classes have more than 100 unique values. I could not find any appropriate encoding possibility. So I created a SQL table, where each value got its ID. Then I extracted the IDs and used it in ML Classification. However the results are poor. So anyone has an idea how to encode such values better?
1 Answer
$\begingroup$
$\endgroup$
Even though you have 100 unique values - Do you see a pattern where perhaps 80 of those unique values appear only once or twice - and the remainder 20 occur much more often.
In that case - you may want to encode the 80 with a special "Other" value. By doing this you lose information - by not being able to distinguish between the 80 categories - but perhaps it doesn't matter to your problem.
The other option is - You can convert your categorical data into an integer (perhaps the ID approach you are talking about ?) and run a Decision Tree model (e.g. Random Forests) on it. My experience with Random Forests is that they work pretty well on such data.
