Feature Importance from Splits

Turn the impurity reductions of the splits in a tree into feature importances. splits is a list of (feature, impurity_reduction) pairs, one per internal node. Sum the reductions per feature and normalise so the importances add up to 1.

Implement solve(splits).

Returns. Return a dict {feature: importance} in order of first appearance. If the total reduction is $00$ the un-normalised sums are returned.

Examples

Example 1

Input:

solve([('a', 1.0), ('b', 2.0), ('a', 1.0)])

Output:

{'a': 0.5, 'b': 0.5}

Example 2

Input:

solve([('x', 0.3), ('y', 0.1)])

Output:

{'x': 0.75, 'y': 0.25}