Reproducibility issue

[janezlapajne]

Hello,

I noticed that results are not reproducible by using the library i.e. when using sklearn drop-down-replacement classes, they will each time produce slightly different results.

For example, when using:

features_engineer = AutoFeatClassifier()
features_engineer.fit_transform(data_train.data, data_train.target.value)

, it will calculate (or select) different features each time.

The issue above I temporarily fixed by using:

 random.seed(seed)
 np.random.seed(seed)

, so that the outputs produced by AutoFeatClassifier stay constant among runs.

However, when I tried using the following:

selector = FeatureSelector(verbose=self.verbose, problem_type="classification", featsel_runs=5)
selector.fit_transform(df_indices, target)

, the above-mentioned seed setting trick didn't translate into desirable outcome - the selected features still change during runs...

Is there an easy fix to correct this? Somewhere in the source randomness must be introduced somewhere, damn.

[janezlapajne]

Also, just now I noticed, that if the number of cores used (acr. n_jobs) is >1, then the results are not reproducible for the first scenario as well. So, the results are reproducible if n_jobs==1 and stochastic if n_jobs==-1.

Cheers.

[jtimko16]

Hey @janezlapajne, thanks for pointing this out. A great observation!

I found several reproducibility issues in the code that I managed to fix. However, there is still some remaining randomness that is unresolved.

For reference the PR in here: (in case you want to review or further contribute)
#45

Cheers,
J.