Prediction explainer for ensemble learning
A method of explaining the reasons for a prediction made by a machine learning ensemble prediction process as to the probability of an outcome for a target observation following training on a plurality of training observations determines the similarity between the target observation and each training observation of a set of said training observations; selects a fraction of the training observations that are most similar to said target observation; ranks the training observations by similarity of each training observation to the target observation; and determines the significance of the features of the ranked training observations to the prediction based upon the increase in variance in a local prediction model when a feature is removed from the local model.
1. A method of explaining the reasons for a prediction made by an machine learning ensemble prediction process as to the probability of an outcome for a target observation, the prediction process having been trained on a plurality of training observations of a training dataset, said target observation and said training observations being characterized by a plurality of features, each feature having a feature value, comprising:
determining similarity between the target observation and each training observation of a set of said training observations;
selecting a subset of said training observations comprising a fraction of said training observations that are most similar to said target observation;
ranking the training observations of said subset by similarity of each such training observation to said target observation; and
determining a significance of the features of said ranked training observations that contributed to said prediction.
2. The method of claim 1 , wherein said determining said similarity comprises determining closeness between the target observation and each training observation.
3. The method of claim 2 , wherein determining said closeness comprises determining a geodesic distance between said target observation and a training observation of said subset.
4. The method of claim 2 , wherein said ensemble prediction process produces a probability of a prediction by combining a plurality of predictions made by a plurality of prediction models of said ensemble prediction process, and said determining said closeness comprises determining a posterior probability for a feature of said target observation relative to a feature of a training observation.
5. The method of claim 4 , wherein said ensemble prediction process comprises a type of Random Forest ensemble process, and said posterior probability comprises the likelihood that said target observation and said training observation are on opposite sides of a split node created by said type of Random Forest ensemble process.
6. The method of claim 1 , wherein said determining said similarity comprises determining a similarity weight for each training observation, said similarity weight measuring closeness of each training observation to said target observation.
7. The method of claim 6 , wherein said method further comprises building a local model by fitting a decision tree the training observations of said subset using said similarity weights.
8. The method of claim 6 , wherein said ranking comprises ranking said training observations of said subset relative to said target observation by similarity weight and to afford separations between classes.
9. The method of claim 1 , wherein said determining the significance of features to said prediction comprises fitting a local model to a second set of weighted training observations to afford a best separation between classes, removing each feature, in turn, from the local model, and determining an increase in variance in the local model upon removing each feature.
10. The method of claim 9 further comprising ordering said features by said increase in variance upon a feature being removed to indicate significance of said feature to said prediction.
11. A computer product comprising a non-transitory computer readable medium storing executable instructions for controlling the operations of a computer to perform a method of explaining the reasons for a prediction made by an machine learning ensemble prediction process as to the probability of an outcome predicted by said prediction process for a target observation, the prediction process having been trained on a plurality of training observations of a training dataset, said target observation and said training observations being characterized by a plurality of features, each feature having a feature value, the method comprising:
determining similarity between the target observation and each training observation of a set of said training observations;
selecting a subset of said training observations comprising a fraction of said training observations that are most similar to said target observation;
ranking the training observations of said subset by similarity of each such training observation to said target observation; and
determining a significance of the features of said ranked training observations that contributed to said prediction.
12. The computer product of claim 11 , wherein said determining said similarity comprises determining closeness between the target observation and each training observation.
13. The computer product of claim 12 , wherein determining said closeness comprises determining a geodesic distance between said target observation and a training observation of said subset.
14. The computer product of claim 12 , wherein said ensemble prediction process produces a prediction by combining a plurality of predictions made by a plurality of prediction models of said ensemble prediction process, and said determining said closeness comprises determining a posterior probability for said target observation relative to a training observation.
15. The computer product of claim 14 , wherein said ensemble prediction process comprises a type of Random Forest ensemble process, and said posterior probability comprises the likelihood that said target observation and said training observation are on opposite sides of a split node created by said type of Random Forest ensemble process.
16. The computer product of claim 11 , wherein said determining said similarity comprises determining a similarity weight for each training observation, said similarity weight measuring closeness of each training observation to said target observation.
17. The computer product of claim 16 , wherein said method further comprises building a local model by fitting a decision tree to said subset of the training observations using said similarity weights.
18. The computer product of claim 16 , wherein said ranking comprises ranking said training observations of said subset relative to said target observation by similarity weight and to afford separations between classes.
19. The computer product of claim 11 , wherein said determining the significance of features to said prediction comprises fitting a local prediction model to a second set of weighted training observations to afford a best separation between classes, removing each feature, in turn, from the local model, and determining an increase in variance in the local model upon removing each feature.
20. The computer product of claim 19 further comprising ordering said features by said increase in variance upon a feature being removed to indicate significance of said feature to said prediction.