A bucket-splitter dynamically resize bucket according to their SSE(Sum of Square Errors).

In each iteration, the bucket with the highest SSE is split into two new buckets, two buckets with the lowest SSE are merged into a new one.

LTD recommended number of iterations is DataSize / threshold * 10 but it depends. For a plot whit one highly fluctuating area and several small peaks, big number of iterations causes small peaks to be lost. So I change the formula to DataSize / threshold / 10 and limit the number to 500.