Fast automatic explanation of scored observations
Systems and methods for generating concise explanations of scored observations that strike good, and computationally efficient, trade-offs between rank-ordering performance and explainability of scored observations are disclosed. The systems and methods described herein for explaining scored observations are based on a framework of partial dependence functions (PDFs), multi-layered neural networks (MNNs), and Latent Explanations Neural Network Scoring (LENNS).
1. A method for improved modeling of automated score observations in a computing system, the method comprising:
modeling a score function for scoring observation of one or more activities in a multilayer neural network (MNN), the MNN comprising processing neurons linked via directed, weighted interconnections, the processing neurons including a set of input neurons for processing input variables to the MNN and at least an output neuron representing at least one score;
restricting connectivity between two or more processing neurons of the MNN to inhibit interactions between at least two or more of the processing neurons in the MNN to produce a plurality of nonlinear univariate functions;
partitioning the input variables to the MNN into at least a first group and a second group;
generating a connectivity graph based on interactions between the set of input neurons in the first group and the second group; and
determining a score for the observation, based on a weighted combination of scores from nonlinear univariate functions corresponding to the set of input neurons in the connectivity graph, to generate one or more explanations for the score.
2. The method in accordance with claim 1 , wherein input variables to the MNN are partitioned into groups according to input representing a meaning of an input variable as provided by a domain expert.
3. The method in accordance with claim 1 , wherein input variables to the MNN are partitioned into groups according to input representing a variable grouping determined by at least one of clustering or principal component analysis on the input variables.
4. The method in accordance with claim 1 , wherein input variables to the MNN are partitioned into groups, the connectivity between two or more processing neurons of the MNN being based on providing a formation of nonlinear group part-scores.
5. The method in accordance with claim 4 , wherein at least one of the nonlinear group part-scores summarizes a signal for a group of input variables to optimize a rank-ordering objective.
6. The method in accordance with claim 4 , wherein a score loss of the scored observation is calculated for at least one of the non-linear group part-scores.
7. The method in accordance with claim 6 , wherein matched dependence values or nearest partial dependence values associated with at least one of the non-linear group part-scores are sorted based on the associated score loss.
8. The method in accordance with claim 1 , wherein a list of explanations associated with the sorted dependence values are generated.
9. The method in accordance with claim 1 , wherein the score function is inhibited from modeling interactions between at least some of the processing neurons via the weighted interconnections.
10. The method in accordance with claim 9 , wherein at least one of the explanations is associated with at least one nonlinear univariate function.
11. A system for improved modeling of automated score observations in a computing system, the system comprising one or more processors for performing the following functions:
modeling a score function for scoring observation of one or more activities in a multilayer neural network (MNN), the MNN comprising processing neurons linked via directed, weighted interconnections, the processing neurons including a set of input neurons for processing input variables to the MNN and at least an output neuron representing at least one score;
restricting connectivity between two or more processing neurons of the MNN to inhibit interactions between at least two or more of the processing neurons in the MNN to produce a plurality of nonlinear univariate functions;
partitioning the input variables to the MNN into at least a first group and a second group;
generating a connectivity graph based on interactions between the set of input neurons in the first group and the second group; and
determining a score for the observation, based on a weighted combination of scores from nonlinear univariate functions corresponding to the set of input neurons in the connectivity graph, to generate one or more explanations for the score.
12. The system in accordance with claim 11 , wherein input variables to the MNN are partitioned into groups according to input representing a meaning of an input variable as provided by a domain expert.
13. The system in accordance with claim 11 , wherein input variables to the MNN are partitioned into groups according to input representing a variable grouping determined by at least one of clustering or principal component analysis on the input variables.
14. The system in accordance with claim 11 , wherein input variables to the MNN are partitioned into groups, the connectivity between two or more processing neurons of the MNN being based on providing a formation of nonlinear group part-scores.
15. The system in accordance with claim 14 , wherein at least one of the nonlinear group part-scores summarizes a signal for a group of input variables to optimize a rank-ordering objective.
16. The system in accordance with claim 14 , wherein a score loss of the scored observation is calculated for at least one of the non-linear group part-scores.
17. The system in accordance with claim 16 , wherein matched dependence values or nearest partial dependence values associated with at least one of the non-linear group part-scores are sorted based on the associated score loss.
18. The system in accordance with claim 11 , wherein a list of explanations associated with the sorted dependence values are generated.
19. The system in accordance with claim 11 , wherein the score function is inhibited from modeling interactions between at least some of the processing neurons via the weighted interconnections.
20. A computer program product comprising a non-transitory machine-readable medium storing instructions that, when executed by at least one programmable processor, cause the at least one programmable processor to perform operations comprising:
modeling a score function for scoring observation of one or more activities in a multilayer neural network (MNN), the MNN comprising processing neurons linked via directed, weighted interconnections, the processing neurons including a set of input neurons for processing input variables to the MNN and at least an output neuron representing at least one score;
restricting connectivity between two or more processing neurons of the MNN to inhibit interactions between at least two or more of the processing neurons in the MNN to produce a plurality of nonlinear univariate functions;
partitioning the input variables to the MNN into at least a first group and a second group;
generating a connectivity graph based on interactions between the set of input neurons in the first group and the second group; and
determining a score for the observation, based on a weighted combination of scores from nonlinear univariate functions corresponding to the set of input neurons in the connectivity graph, to generate one or more explanations for the score.