System and method for estimating performance of a classifier
View Patent ↗A method for estimating the performance of a statistical classifier. The method includes inputting a first set of business data in a first format from a real business process and storing the first set of business data in the first format into memory. The method applying a statistical classifier to the first set of business data and recording its classification decisions and obtaining a labeling that contains the correct decision for each data item. The method includes computing a weight for each data item that reflects its true frequency and computing a performance measure of the statistical classifier based on the weights that reflect true frequency. The method also displays the performance measure to a user.
1. A method for estimating the performance of a statistical classifier, the method comprising:
inputting a first set of business data in a first format from a real business process;
storing the first set of business data in the first format into memory;
applying a statistical classifier to the first set of business data;
recording classification decisions from the statistical classifier based upon the first data set;
obtaining a labeling that contains a true classification decision for each data item from the first set of business data;
computing a performance measure of the statistical classifier based upon the labeling that contains a true classification decision for each data item from the first set of business data;
computing a weight for each data item that reflects its true frequency;
correcting the performance measure of the statistical classifier based on the weights that reflect true frequency;
displaying the corrected performance measure to a user.
2. The method in 1 wherein the labeling is obtained from a domain expert.
3. The method in 1 wherein the first data set is created through active learning.
4. The method in 1 wherein the performance measure is one of precision, recall or a combination of precision and recall.
5. The method in 4 wherein the combination is a weighted sum.
6. The method in 1 wherein the first data set comprises text.
7. The method in 1 wherein the true weight of a data item is computed using a second data set drawn randomly from the population.
8. The method in 7 wherein the true weight of a data item is computed as a ratio of the biased weight of the data item in the second data set divided by the unbiased weight of the data item in the first data set.
9. The method in 8 wherein the biased weight is computed by one of histogram based method, kernel based method or expansion to basis function method.
10. The method in 8 wherein the unbiased weight is computed by one of histogram based method, kernel based method or expansion to basis function method.
11. The method in 7 wherein the first data set and the second data set are disjoint.
12. The method in 1 wherein the performance measure is computed by replacing each data item's contribution to the performance measure by the product of its true weight and its contribution.
13. The method in 1 wherein the performance measure is computed as a combination of the performance measure with the use of true weights and the performance measure without the use of true weights.
14. A method for estimating the performance of a statistical classifier, the method comprising:
inputting a first set of business data in a first format from a real business process;
storing the first set of business data in the first format into memory;
applying a statistical classifier to the first set of business data and recording its classification decisions;
obtaining a labeling that contains the correct decision for each data item;
computing a weight for each data item that reflects its true frequency;
computing a performance measure of the statistical classifier based on the weights that reflect true frequency;
displaying the performance measure to a user.
15. The method in 14 wherein the labeling is obtained from a domain expert.
16. The method in 14 wherein the first data set is created through active learning.
17. The method in 14 wherein the performance measure is one of precision, recall or a combination of precision and recall.
18. The method in 14 wherein the first data set comprises text.
19. The method in 14 wherein the true weight of a data item is computed using a second data set drawn randomly from the population.
20. The method in 19 wherein the true weight of a data item is computed as a ratio of the biased weight of the data item in the second data set divided by the unbiased weight of the data item in the first data set.
21. The method in 20 wherein the biased weight is computed by one of histogram based method, kernel based method or expansion to basis function method.
22. The method in 20 wherein the unbiased weight is computed by one of histogram based method, kernel based method or expansion to basis function method.
23. The method in 19 wherein the first data set and the second data set are disjoint.
24. The method in 14 wherein the performance measure is computed by replacing each data item's contribution to the performance measure by the product of its true weight and its contribution.
25. The method in 14 wherein the performance measure is computed as a combination of the performance measure with the use of true weights and the performance measure without the use of true weights.
26. The method in 25 wherein the combination is a weighted sum.
27. A system for estimating the performance of a statistical classifier, the system comprising one or more memories, the one or more memories including:
one or more codes for receiving a first set of business data in a first format from a real business process;
one or more codes for storing the first set of business data in the first format into memory;
one or more codes for applying a statistical classifier to the first set of business data;
one or more codes for recording classification decisions from the statistical classifier based upon the first data set;
one or more codes for obtaining a labeling that contains a true classification decision for each data item from the first set of business data;
one or more codes for computing a performance measure of the statistical classifier based upon the labeling that contains a true classification decision for each data item from the first set of business data;
one or more codes for computing a weight for each data item that reflects its true frequency;
one or more codes for correcting the performance measure of the statistical classifier based on the weights that reflect true frequency; and
one or more codes for displaying the corrected performance measure to a user.
28. A system for estimating the performance of a statistical classifier, the system comprising one or more memories, the one or more memories including:
means for receiving a first set of business data in a first format from a real business process;
means for storing the first set of business data in the first format into memory;
means for applying a statistical classifier to the first set of business data;
means for recording classification decisions from the statistical classifier based upon the first data set;
means for obtaining a labeling that contains a true classification decision for each data item from the first set of business data;
means for computing a performance measure of the statistical classifier based upon the labeling that contains a true classification decision for each data item from the first set of business data;
means for computing a weight for each data item that reflects its true frequency;
means for correcting the performance measure of the statistical classifier based on the weights that reflect true frequency; and
means for displaying the corrected performance measure to a user.