How to combine different results of a stochastic classifier I'm writing a paper about a machine learning-based system and using CNNs on a GPU cluster to compare two methods of feature engineering. Due to the non-deterministic nature of the algorithm, I couldn't get the same results on every run. For getting trustable results I run tests ten times for each method but, I don't know how to aggregate ten accuracy, precision, recall and f-score value to one. Is there a standard way to aggregate these scores, or can I just average these values?
 A: I ran into the same issue in a similar study/publication of mine. Rather than "aggregate to one", I would suggest maybe producing a graphic/plotting summary to illustrate the comparison of the two methods of feature-engineering. 
For precision and recall, since people are used to seeing precision-recall curves, you can combine the two to produce a precision-recall "confidence region" plot (note, "confidence" is used in a very casual and not statistical way) as below: 

and for individual accuracy measures, I'd say boxplots are better than averages because you can show how "stable" the runs are by showing whether the boxplots are tight or wide (like I did here)

Here's the citation and biorxiv link to the publication (I believe BMC Medical Genomics might be behind a paywall). 
Rachid Zaim, S., Kenost, C., Berghout, J., Vitali, F., Zhang, H. H., & Lussier, Y. A. (2019). Evaluating single-subject study methods for personal transcriptomic interpretations to advance precision medicine. BMC Medical Genomics, 12(5).
https://www.biorxiv.org/content/10.1101/428581v1
