I'm a student studying a data mining course and have come across a problem.
I need to explain the problem with the help of an example scenario as I do not know how to explain the problem in any other way.
I have n 'manufacture dates' (classes). On a particular day, y identical balls are to be segregated into these n classes based on their date of manufacture. I have x identical balls left for which I do not know the dates of manufacture (unclassified), i.e. y-x balls were classified.
Let us call the mean/standard deviation found when all balls were classified correctly to be the "true" mean/standard deviation (i.e. mean/standard deviation when all y balls are correctly classified)
Since we do not know the classes the x balls belong to, how can we distribute the x balls amongst the n classes so that we get a mean/standard deviation that is close to the true mean/standard deviation ?
Case1: Value of x (i.e. number of balls not classified on a particular day) is relatively small to the total number of balls given on the same day (maybe 1 out of 10 balls are not classified).
In this case, I considered ignoring the balls that haven't been classified as the difference between the final result (mean and standard deviation of the balls classified) and the true mean/standard deviation would be minute.
Case2: Value of x is relatively large (i.e 4 out of 10 balls are not classified OR 6 out of 10 balls are not classified).
In this case, I do not prefer ignoring the unclassified balls as the end result (mean, standard deviation) will be erroneous. Rather, I thought of performing an even distribution or weighted distribution of the unclassified balls into the classes, but I do not have good reasoning for why this would be a good approach other than the point of not ignoring majority of the balls.