Systems and methods for classifying imbalanced data
A transaction classification system obtains a data set including first data associated with a first characteristic and second data associated with a second characteristic. In response to obtaining the data set, the system uses a classification model to generate a classification by classifying the first data into majority data and the second data into minority data. From the classification and using the classification model, the system determines a cost. The system modifies the classification model based on this cost to generate an updated classification model. The system uses the updated classification model to re-classify a subset of the first data into the minority data and a subset of the second data into the majority data.
1. A computer-implemented method comprising:
receiving, by a computer, a request to obtain a classification model that satisfies one or more criteria, wherein the one or more criteria include a quantifiable loss below a threshold value;
obtaining, by the computer, an imbalanced data set including first data associated with a first characteristic and second data associated with a second characteristic, wherein the imbalanced data set is obtained from a repository of data sets usable for training machine learning models;
generating, by the computer, a modified loss function to calculate the quantifiable loss, wherein the modified loss function is generated by modifying an original loss function to include a weighing component corresponding to a loss resulting from false positive and false negative classifications, and wherein the original loss function generates an esoteric statistical quantity;
training, by the computer, an initial iteration of the classification model to generate a classification, wherein the initial iteration is trained by randomly initializing a set of model coefficients corresponding to an input space of the imbalanced data set;
generating, by the computer, an initial quantifiable loss corresponding to the initial iteration, wherein the initial quantifiable loss is generated using the modified loss function, and wherein the initial quantifiable loss results from a misclassification of the imbalanced data set;
iteratively training, by the computer, the classification model until a modified classification model is obtained that satisfies the one or more criteria, wherein the classification model is iteratively trained as a result of corresponding quantifiable losses failing to satisfy the one or more criteria, and wherein iteratively training the classification model includes updating the set of model coefficients according to the corresponding quantifiable losses; and
providing, by the computer, the modified classification model, wherein the modified classification model is provided as a result of a quantifiable loss corresponding to the modified classification model being below the threshold value.
2. The computer-implemented method of claim 1 , further comprising:
obtaining a new imbalanced data set; and
classifying data points in the new imbalanced data set as being associated with either the first characteristic or the second characteristic, wherein classifying the data points is performed using the modified classification model.
3. The computer-implemented method of claim 1 , wherein:
the first characteristic corresponds to fraudulent transactions processed by a point-of-sale system; and
the second characteristic corresponds to authenticated transactions processed by the point-of-sale system.
4. The computer-implemented method of claim 1 , wherein the set of model coefficients are updated using a gradient descent algorithm.
5. The computer-implemented method of claim 1 , further comprising:
providing output of the modified classification model, wherein the output includes data points organized according to individual associations with the first characteristic and the second characteristic.
6. The computer-implemented method of claim 1 , wherein the set of model coefficients are randomly initialized according to a Gaussian distribution with a variance centered around zero.
7. The computer-implemented method of claim 1 , wherein the set of model coefficients are updated subject to a pre-determined learning rate.
8. The computer-implemented method of claim 1 , wherein the classification model is a logistic regression model, and wherein the logistic regression model generates probabilities that data points have a particular classification.
9. The computer-implemented method of claim 1 , wherein imbalanced data set includes a set of data points corresponding to simulated transactions.
10. The computer-implemented method of claim 1 , wherein the quantifiable losses correspond to monetary costs resulting from misclassifications of the imbalanced data set.
11. A system, including:
one or more processors; and
memory including instructions that, as a result of being executed by the one or more processors, cause the system to:
receive a request to obtain a classification model that satisfies one or more criteria, wherein the one or more criteria include a quantifiable loss below a threshold value;
obtain an imbalanced data set including first data associated with a first characteristic and second data associated with a second characteristic, wherein the imbalanced data set is obtained from a repository of data sets usable for training machine learning models;
generate a modified loss function to calculate the quantifiable loss, wherein the modified loss function is generated by modifying an original loss function to include a weighing component corresponding to a loss resulting from false positive and false negative classifications, and wherein the original loss function generates an esoteric statistical quantity;
train an initial iteration of the classification model to generate a classification, wherein the initial iteration is trained by randomly initializing a set of model coefficients corresponding to an input space of the imbalanced data set;
generate an initial quantifiable loss corresponding to the initial iteration, wherein the initial quantifiable loss is generated using the modified loss function, and wherein the initial quantifiable loss results from a misclassification of the imbalanced data set;
iteratively train the classification model until a modified classification model is obtained that satisfies the one or more criteria, wherein the classification model is iteratively trained as a result of corresponding quantifiable losses failing to satisfy the one or more criteria, and wherein iteratively training the classification model includes updating the set of model coefficients according to the corresponding quantifiable losses; and
provide the modified classification model, wherein the modified classification model is provided as a result of a quantifiable loss corresponding to the modified classification model being below the threshold value.
12. The system of claim 11 , wherein the instructions further cause the system to:
obtain a new imbalanced data set; and
use the modified classification model to classify data points in the new imbalanced data set as being associated with either the first characteristic or the second characteristic.
13. The system of claim 11 , wherein:
the first characteristic corresponds to fraudulent transactions processed by a point-of-sale system; and
the second characteristic corresponds to authenticated transactions processed by the point-of-sale system.
14. The system of claim 11 , wherein the set of model coefficients are updated using a gradient descent algorithm.
15. The system of claim 11 , wherein the instructions further cause the system to:
provide output of the modified classification model, wherein the output includes data points organized according to individual associations with the first characteristic and the second characteristic.
16. The system of claim 11 , wherein the set of model coefficients are randomly initialized according to a Gaussian distribution with a variance centered around zero.
17. The system of claim 11 , wherein the set of model coefficients are updated subject to a pre-determined learning rate.
18. The system of claim 11 , wherein the classification model is a logistic regression model, and wherein the logistic regression model generates probabilities that data points have a particular classification.
19. The system of claim 11 , wherein imbalanced data set includes a set of data points corresponding to simulated transactions.
20. The system of claim 11 , wherein the quantifiable losses correspond to monetary costs resulting from misclassifications of the imbalanced data set.
21. A non-transitory computer-readable storage medium storing thereon executable instructions that, as a result of being executed by one or more processors of a computer system, cause the computer system to:
receive a request to obtain a classification model that satisfies one or more criteria, wherein the one or more criteria include a quantifiable loss below a threshold value;
obtain an imbalanced data set including first data associated with a first characteristic and second data associated with a second characteristic, wherein the imbalanced data set is obtained from a repository of data sets usable for training machine learning models;
generate a modified loss function to calculate the quantifiable loss, wherein the modified loss function is generated by modifying an original loss function to include a weighing component corresponding to a loss resulting from false positive and false negative classifications, and wherein the original loss function generates an esoteric statistical quantity;
train an initial iteration of the classification model to generate a classification, wherein the initial iteration is trained by randomly initializing a set of model coefficients corresponding to an input space of the imbalanced data set;
generate an initial quantifiable loss corresponding to the initial iteration, wherein the initial quantifiable loss is generated using the modified loss function, and wherein the initial quantifiable loss results from a misclassification of the imbalanced data set;
iteratively train the classification model until a modified classification model is obtained that satisfies the one or more criteria, wherein the classification model is iteratively trained as a result of corresponding quantifiable losses failing to satisfy the one or more criteria, and wherein iteratively training the classification model includes updating the set of model coefficients according to the corresponding quantifiable losses; and
provide the modified classification model, wherein the modified classification model is provided as a result of a quantifiable loss corresponding to the modified classification model being below the threshold value.
22. The non-transitory computer-readable storage medium of claim 21 , wherein the executable instructions further cause the computer system to:
obtain a new imbalanced data set; and
use the modified classification model to classify data points in the new imbalanced data set as being associated with either the first characteristic or the second characteristic.
23. The non-transitory computer-readable storage medium of claim 21 , wherein:
the first characteristic corresponds to fraudulent transactions processed by a point-of-sale system; and
the second characteristic corresponds to authenticated transactions processed by the point-of-sale system.
24. The non-transitory computer-readable storage medium of claim 21 , wherein the set of model coefficients are updated using a gradient descent algorithm.
25. The non-transitory computer-readable storage medium of claim 21 , wherein the executable instructions further cause the computer system to:
provide output of the modified classification model, wherein the output includes data points organized according to individual associations with the first characteristic and the second characteristic.
26. The non-transitory computer-readable storage medium of claim 21 , wherein the set of model coefficients are randomly initialized according to a Gaussian distribution with a variance centered around zero.
27. The non-transitory computer-readable storage medium of claim 21 , wherein the set of model coefficients are updated subject to a pre-determined learning rate.
28. The non-transitory computer-readable storage medium of claim 21 , wherein the classification model is a logistic regression model, and wherein the logistic regression model generates probabilities that data points have a particular classification.
29. The non-transitory computer-readable storage medium of claim 21 , wherein imbalanced data set includes a set of data points corresponding to simulated transactions.
30. The non-transitory computer-readable storage medium of claim 21 , wherein the quantifiable losses correspond to monetary costs resulting from misclassifications of the imbalanced data set.