Scoring a population of examples using a model
A computer-implemented method for scoring an example with causal information includes a computer system receiving an indication of a goal and applying a predictive model to the example to yield an output score with respect to the goal. The computer system produces causal score for each feature-value pair in the example, each causal score indicating of a relative extent to which the feature-value pair is responsible for influencing the output score. Then, the computer system presents the output score and the causal score for each feature-value pair in the example.
1. A computer-implemented method for scoring a population of examples subject to an additional constraint of a fixed aggregate resource expenditure, the method comprising:
generating a master gain ratio list that identifies, for each example in the population of examples, an attribute combination having a largest gain ratio among all attribute combinations in the example, wherein the master gain ratio list is generated using an iterative process comprising:
identifying a plurality of attribute combinations associated with the example, wherein for each attribute combination of the plurality of attribute combinations at least one attribute in the attribute combination indicates a resource expenditure for the attribute combination,
using a predictive model to determine a predicted gain value for each attribute combination,
determining a gain ratio for each attribute combination, wherein the gain ratio for an attribute combination is defined as the predicted gain value for the attribute combination divided by the resource expenditure for the attribute combination;
identifying a particular attribute combination in the plurality of attribute combinations having the largest gain ratio, and
adding the particular attribute combination and the largest gain ratio to the master gain ratio list;
sorting, by the computer system, the master gain ratio list according to the gain ratio; and
performing the resource expenditure associated with a first attribute combination on the sorted master gain ratio list.
2. The method of claim 1 , further comprising:
after the iterative process is complete, removing any item from the master gain ratio list corresponding to a resource expenditure exceeding a predetermined constraint.
3. The method of claim 1 , wherein the population of examples comprises a plurality of continuous feature values and the method comprises:
binning the plurality of continuous feature values into equal intervals, yielding a plurality of bins,
wherein the iterative process is applied to each of the plurality of bins.
4. The method of claim 1 , wherein the iterative process further comprises:
for each attribute combination of the plurality of attribute combinations, adjusting the gain ratio based on a monetary cost associated with performing the resource expenditure indicated by the attribute combination.
5. The method of claim 1 , wherein the iterative process further comprises:
determining a cost for each individual value associated with feature-value pairs included in the example; and
for each of the plurality of attribute combinations, adjusting the gain ratio based on the cost corresponding to values associated with feature-value pairs in the attribute combination.
6. The method of claim 1 , wherein the iterative process further comprises:
for each of the plurality of attribute combinations, using a cost function associated with achieving and not achieving a goal to determine a goal probability estimation; and
adjusting the gain ratio associated with each attribute combination based on the corresponding goal probability estimation.
7. A system for scoring a population of examples subject to an additional constraint of a fixed aggregate resource expenditure, the system comprising:
one or more processors; and
a non-transitory, computer-readable storage medium in operable communication with the one or more processors, wherein the computer-readable storage medium contains one or more programming instructions that, when executed, cause the processors to:
generate a master gain ratio list that identifies, for each example in the population of examples, an attribute combination having a largest gain ratio among all attribute combinations in the example, wherein the master gain ratio list is generated-using an iterative process comprising:
identifying a plurality of attribute combinations associated with the example, wherein for each attribute combination of the plurality of attribute combinations at least one attribute in the attribute combination indicates a resource expenditure for the attribute combination,
using a predictive model to determine a predicted gain value for each attribute combination,
determining a gain ratio for each attribute combination, wherein the gain ratio for an attribute combination is defined as the predicted gain value for the attribute combination divided by the resource expenditure for the attribute combination,
identifying a particular attribute combination in the plurality of attribute combinations having a largest gain ratio, and
adding the particular attribute combination and the largest gain ratio to the master gain ratio list;
sort the master gain ratio list according to the gain ratio; and
perform the resource expenditure associated with a first attribute combination on the sorted master gain ratio list.
8. The system of claim 7 , further comprising:
after the iterative process is complete, removing any item from the master gain ratio list corresponding to a resource expenditure exceeding a predetermined constraint.
9. The system of claim 7 , wherein the population of examples comprises a plurality of continuous feature values and the system comprises:
binning the plurality of continuous feature values into equal intervals, yielding a plurality of bins,
wherein the iterative process is applied to each of the plurality of bins.
10. The system of claim 7 , wherein the iterative process further comprises:
for each attribute combination of the plurality of attribute combinations, adjusting the gain ratio based on a monetary cost associated with performing the resource expenditure indicated by the attribute combination.
11. The system of claim 7 , wherein the iterative process further comprises:
determining a cost for each individual value associated with feature-value pairs included in the example; and
for each of the plurality of attribute combinations, adjusting the gain ratio based on the cost corresponding to values associated with feature-value pairs in the attribute combination.
12. The system of claim 7 , wherein the iterative process further comprises:
for each of the plurality of attribute combinations, using a cost function associated with achieving and not achieving a goal to determine a goal probability estimation; and
adjusting the gain ratio associated with each attribute combination based on the corresponding goal probability estimation.