The FPGrowth algorithm should always return the same result as Apriori. In SPMF, I have tested the algorithms very well, and all 11 algorithms for frequent itemset mining (Apriori, FPGrowth, LCM, HMine, Relim, PrePost, Fin, etc.) returns the same result, as it should.
For Weka, I don't use Weka, so I cannot tell you why the results are incorrect for their implementation. Maybe their implementation has a bug. Or maybe that it does not handle the parameters as it should. For example, to find frequent patterns, maybe they use > minsup instead of >= minsup... I don't know. I had a look at their implementation a few years ago, and their implementation is not very good. Actually, Weka is quite slow in general. I think that they did not optimize much their implementation. Here are some experimental comparison that I did a few years ago that shows that their implementation is quite slow and consumes a lot of memory (and this was before I made some considerable additional optimizations to further improve the performance in SPMF):
By the way, from what I have heard a year ago, the mapreduce implementation of FPGrowth in Mahout also had some bugs.
Edited 6 time(s). Last edit at 01/25/2017 01:12AM by webmasterphilfv.