In frequency set mining, for large sets, we can use sample to reduce the calculation while not miss too much for the support using formula:
n > -2 * log(ci)/ (supp * epsilon^2) #with ci = 1 - confidence level supp = the lowest support you request for this transaction set epsilon = the error rate between the real support and support calculated with sample
as long as we make sure n is greater than the formula, then we has 1-ci confidence level that the sample support is within 1-epsilon support
Is there some formular similar for lift and confidence?
In short, Is there a way to sample the transaction set while keep the lift and/or confidence of each rule remains not too far away from the original set?