Avoid quadratic loop in getNumericAttributeEntropy.
We don't need to recalculate whole distribution for each split,
just move changed values. Also use array of slices instead of
map of maps of strings to avoid map overhead.
For our case I see time reductions from 100+ hours to 50 minutes.
I've added benchmark with synthetic data (iris.csv repeated 100 times)
and it also shows a nice improvement:
name old time/op new time/op delta
RandomForestFit-8 117s ± 4% 0s ± 1% -99.61% (p=0.001 n=5+10)
0 is a rounding quirk of benchstat, it should be closer to 0.5s:
name time/op
RandomForestFit-8 460ms ± 1%
This patch adds:
* Gini index and information gain ratio as
DecisionTree split options;
* handling for numeric Attributes (split point
chosen naïvely on the basis of maximum entropy);
* A couple of additional utility functions in base/
* A new dataset (see sources.txt) for testing.
Performance on Iris performs markedly without discretisation.
