2016-ml-contest

2016-ml-contest copied to clipboard

HI @kwinkunks, I know it has been one year already, I just happen to take a look at this repo again and found in the utils.py that the score used is "accuracy", not the actual "F1 score", is that right?

Hi... Indeed, I used the accuracy function in utils.py. This is the same as sklearn.metrics.f1_score with average='micro'.

This was discussed in another issue. It seems that using average='weighted' may have been more sensible for a dataset with such imbalanced labels, but 'micro' is what we went with.

It would be interesting to test them against each other, because one question I have is whether weighting the small populations would make the results rather unstable, as single instances of small populations could substantially change the score. Since I was using 100 realizations of the scores for the final ranking, maybe this would not have been a big concern, but it's worth investigating I think.

I see, I do remember that I saw this discussion somewhere, but couldn't find it last night. I agree that average='weighted would be a better choice for such imbalanced data set, since 'micro' basically gives the accuracy in normal sense.

If some population is too small , I guess their precision and recall would not be weighted too much towards the total.