Method for configuring multiple machine learning classifiers
Configuring a multi-classification system having multiple component classifiers includes storing of data records that represent different levels of performance of the system. The component classifiers are configured with corresponding decision threshold values contained in a selected one of the data records. Performance of the multi-classification system subsequent to configuring the component classifiers is approximated by the performance level associated with the selected data record.
1. A method in a system comprising a plurality of component classifiers, the method comprising:
storing a plurality of data records that represent measures of performance of the system, each data record associated with a performance level of the system, each data record comprising:
a plurality of decision thresholds for the plurality of component classifiers;
a system true-positive rate (TPR) corresponding to the plurality of decision thresholds in said each data record; and
a system false-positive rate (FPR) corresponding to the plurality of decision thresholds in said each data record,
the performance level associated with said each data record being based on the system TPR and the system FPR;
configuring the plurality of component classifiers comprising the system with the plurality of decision thresholds in a data record selected from among the plurality of data records based on the performance levels of the data records; and
operating the system according to the configured plurality of component classifiers, wherein the system operates at a performance level that is approximated by the performance level associated with the selected data record.
2. The method of claim 1 , further comprising presenting the plurality of data records to a user and receiving information from the user indicative of one of the plurality of data records as the selected data record.
3. The method of claim 1 , wherein the plurality of data records is a subset of a plurality of candidate data records that is greater in number than the plurality of data records.
4. The method of claim 1 , wherein the plurality of data records represents a Pareto frontier among a plurality of candidate data records that is greater in number than the plurality of data records.
5. The method of claim 1 , wherein the system TPR and system FPR in a given data record characterizes the system when the plurality of component classifiers comprising the system are configured with the plurality of decision thresholds in the given data record.
6. The method of claim 5 , wherein the system TPR and system FPR that characterize the system are determined by:
presenting the plurality of component classifiers comprising the system with corresponding training data sets to determine component TPRs and FPRs of the component classifiers; and
computing the system TPR and system FPR from the component TPRs and FPRs.
7. The method of claim 1 , wherein each system FPR is based at least on a component TPR and a component FPR of a first component classifier comprising the system and on a component TPR and a component FPR of a second component classifier comprising the system.
8. The method of claim 7 , wherein each system TPR is based on only the component TPR of a first component classifier comprising the system and on only the component FPR of a second component classifier comprising the system.
9. The method of claim 1 , wherein the plurality of component classifiers includes an image classifier to classify machine-generated images and a report classifier to classify human-originated reports.
10. The method of claim 9 , wherein the system FPR in a given data record is based on a TPR and an FPR of the image classifier configured with a first decision threshold among the plurality of decision thresholds in the given data record and a TPR and an FPR of the report classifier configured with a second decision threshold among the plurality of decision thresholds in the given data record.
11. The method of claim 10 , wherein the system FPR is further based on base rates that are independent of performance of the image classifier and performance of the report classifier.
12. A system comprising a plurality of component classifiers and means for configuring the plurality of component classifiers comprising:
means for storing a plurality of data records that represent measures of performance of the system, each data record associated with a performance level of the system, each data record comprising:
a plurality of decision thresholds for the plurality of component classifiers;
a system true-positive rate (TPR) corresponding to the plurality of decision thresholds in said each data record; and
a system false-positive rate (FPR) corresponding to the plurality of decision thresholds in said each data record,
the performance level associated with said each data record being based on the system TPR and the system FPR; and
means for configuring the plurality of component classifiers comprising the system with the plurality of decision thresholds in a data record selected from among the plurality of data records based on the performance levels of the data records,
wherein when operating the system according to the configured plurality of component classifiers, the system operates at a performance level that is approximated by the performance level associated with the selected data record.
13. The system of claim 12 , further comprising means for presenting the plurality of data records to a user and receiving information from the user indicative of one of the plurality of data records as the selected data record.
14. The system of claim 12 , wherein the plurality of data records is a subset of a plurality of candidate data records that is greater in number than the plurality of data records.
15. The system of claim 12 , wherein the plurality of data records represents a Pareto frontier among a plurality of candidate data records that is greater in number than the plurality of data records.
16. The system of claim 12 , wherein the system TPR and system FPR in a given data record characterizes the system when the plurality of component classifiers comprising the system are configured with the plurality of decision thresholds in the given data record.
17. The system of claim 16 , wherein the system TPR and system FPR that characterize the system are determined by:
presenting the plurality of component classifiers comprising the system with corresponding training data sets to determine component TPRs and FPRs of the component classifiers; and
computing the system TPR and system FPR from the component TPRs and FPRs.
18. A non-transitory computer-readable storage medium in a system comprising a plurality of component classifiers, the non-transitory computer-readable storage medium having stored thereon computer executable instructions, which when executed by a computer device, cause the computer device to:
store a plurality of data records that represent measures of performance of the system, each data record associated with a performance level of the system, each data record comprising:
a plurality of decision thresholds for the plurality of component classifiers;
a system true-positive rate (TPR) corresponding to the plurality of decision thresholds in said each data record; and
a system false-positive rate (FPR) corresponding to the plurality of decision thresholds in said each data record,
the performance level associated with said each data record being based on the system TPR and the system FPR;
configure the plurality of component classifiers comprising the system with the plurality of decision thresholds in a data record selected from among the plurality of data records based on the performance levels of the data records; and
operate the system according to the configured plurality of component classifiers, wherein the system operates at a performance level that is approximated by the performance level associated with the selected data record.
19. The non-transitory computer-readable storage medium of claim 18 , wherein the computer executable instructions, which when executed by the computer device, further cause the computer device to present the plurality of data records to a user and receiving information from the user indicative of one of the plurality of data records as the selected data record.
20. The non-transitory computer-readable storage medium of claim 18 , wherein the plurality of data records is a subset of a plurality of candidate data records that is greater in number than the plurality of data records.
21. The non-transitory computer-readable storage medium of claim 18 , wherein the plurality of data records represents a Pareto frontier among a plurality of candidate data records that is greater in number than the plurality of data records.
22. The non-transitory computer-readable storage medium of claim 18 , wherein the system TPR and system FPR in a given data record characterizes the system when the plurality of component classifiers comprising the system are configured with the plurality of decision thresholds in the given data record.
23. An apparatus in a system comprising a plurality of component classifiers, the apparatus comprising:
one or more computer processors; and
a computer-readable storage medium comprising instructions for controlling the one or more computer processors to be operable to:
store a plurality of data records that represent measures of performance of the system, each data record associated with a performance level of the system, each data record comprising:
a plurality of decision thresholds for the plurality of component classifiers;
a system true-positive rate (TPR) corresponding to the plurality of decision thresholds in said each data record; and
a system false-positive rate (FPR) corresponding to the plurality of decision thresholds in said each data record,
the performance level associated with said each data record being based on the system TPR and the system FPR;
configure the plurality of component classifiers comprising the system with the plurality of decision thresholds in a data record selected from among the plurality of data records based on the performance levels of the data records; and
operate the system according to the configured plurality of component classifiers, wherein the system operates at a performance level that is approximated by the performance level associated with the selected data record.
24. The apparatus of claim 23 , wherein the computer-readable storage medium further comprises instructions for controlling the one or more computer processors to be operable to present the plurality of data records to a user and receiving information from the user indicative of one of the plurality of data records as the selected data record.
25. The apparatus of claim 23 , wherein the plurality of data records is a subset of a plurality of candidate data records that is greater in number than the plurality of data records.
26. The apparatus of claim 23 , wherein the plurality of data records represents a Pareto frontier among a plurality of candidate data records that is greater in number than the plurality of data records.
27. The apparatus of claim 23 , wherein the system TPR and system FPR in a given data record characterizes the system when the plurality of component classifiers comprising the system are configured with the plurality of decision thresholds in the given data record.