- Training set size: 5200
- Test set size: 1000
I'm currently working on a classifier which classifies images in a class (binary). My classifier uses a bagging approach (10 bags). I've been experimenting a lot and tried to look at the normal bagging vs something I've tried myself for computational reasons: instead of fitting the model in each iteration on a set as large as the training set, I limited the size of the set to 2000. Now the result is that the definite model of the normal bagging has 10% lower classification accuracy than the model trained on the bags of size 2000. I don't see an explanation to this, since my approach uses smaller sets to train on so I'd assume it produces a classifier with worse accuracy.
Note: It's not a random result, I ran both algorithms multiple times and the approach with sets of size 2000 consistently gives a better prediction accuracy.