Hi, I noticed a bad precision of FMeasure class. I think the issue is that the current implementation is summing divisions. It computes the precision and recall for every sample, and after adds it all to compute the overall result. By doing that, the error related to each division are summed and can impact the final result. I tried a new implementation that can fix it. If everyone agrees I'll create an issue at Jira to commit the modified version.
Regards William
