Method of machine learning training for data augmentation
A method including training, using accepted data having known labels, an untrained prediction model to generate a trained prediction model. The method also includes generating, using the trained prediction model, predicted labels for unknown data. The method also includes generating, using a confidence model, confidence scores for the predicted labels. The method also includes selecting, using the confidence scores, a subset of the unknown data and confident labels. The confident labels include ones of the predicted labels for the subset of the unknown data. The method also includes training, using the subset of the unknown data and the confident labels, the trained prediction model to generate an updated prediction model.
1 . A method for self-training a prediction machine learning model by increasing a size of initially available training data, the method comprising:
training, using accepted data having known labels, an untrained supervised prediction model to generate a trained prediction model, wherein:
the accepted data comprises less than a predetermined amount of training data to achieve a desired accuracy of the trained prediction model,
the accepted data comprises the initially available training data, and
the untrained supervised prediction model comprises first parameters;
generating, using the trained prediction model, predicted labels for unknown data;
generating, using a confidence model, confidence scores for the predicted labels, wherein the confidence model comprises a logistic regression algorithm comprising second parameters different than the first parameters;
selecting, using the confidence scores, a subset of the unknown data and confident labels, the confident labels comprising ones of the predicted labels for the subset of the unknown data, wherein selecting further comprises:
selecting, as the confident labels, ones of the predicted labels having predicted label values that exceed a threshold value, and
selecting, as the subset of the unknown data, portions of the unknown data corresponding to the confident labels,
wherein the subset of unknown data is less than the unknown data;
training, using the accepted data, the subset of the unknown data, and the confident labels, the trained prediction model to generate an updated supervised prediction model, wherein the subset of the unknown data, the confident labels, and the accepted data together comprise augmented data that, when used to train the trained prediction model, improves an accuracy of predictions made by the updated supervised prediction model;
executing the updated supervised prediction model on remaining data to generate updated prediction labels, wherein the remaining data comprises the unknown data less the subset of the unknown data;
executing the confidence model on the remaining data to output other confidence scores for the updated predicted labels;
selecting other confident labels for a second subset of the remaining data, outside the subset of data, wherein:
the other confident labels comprise ones of the updated predicted labels for the second subset for which the other confident scores exceed a second threshold, and
the other confident labels and the second subset of the remaining data together increase the size of the initially available training data to increased training data;
training, using the second subset and the other confident labels, the updated supervised prediction model to generate a deployed prediction model, wherein the deployed prediction model is more accurate than the trained prediction model trained using the initially available training data due to the increased training data; and
deploying the deployed prediction model to an enterprise environment.
2 . The method of claim 1 further comprising:
iteratively generating the updated prediction labels, generating the confidence scores, selecting the subset of unknown data, and training the updated supervised prediction model, until a stop condition occurs.
3 . The method of claim 2 , wherein the stop condition comprises a pre-determined number of iterations.
4 . The method of claim 2 , wherein the stop condition comprises finding less than a remainder threshold number of updated confidence scores that are above the threshold value.
5 . The method of claim 1 , further comprising:
receiving live data; and
predicting, using the updated supervised prediction model, live prediction labels for the live data.
6 . The method of claim 5 , wherein the live prediction labels comprise an evaluation of credit risk for a plurality of users from whom the live data was received.
7 . The method of claim 1 , wherein the untrained supervised prediction model comprises a gradient boosted tree algorithm.
8 . A system for self-training a deployed prediction model by increasing a size of initially available training data, the system comprising:
a computer processor;
a data repository in communication with the computer processor;
a training application executable on the computer processor, wherein the training application comprises program code which, when executed, performs a computer-implemented method comprising:
training, using accepted data having known labels, an untrained supervised prediction model to generate a trained prediction model, wherein:
the accepted data comprises less than a predetermined amount of training data to achieve a desired accuracy of the trained prediction model, the accepted data comprises the initially available training data, and
the untrained supervised prediction model comprises first parameters;
generating, using the trained prediction model, predicted labels for unknown data;
generating, using a confidence model, confidence scores for the predicted labels, wherein the confidence model comprises a logistic regression algorithm comprising second parameters different than the first parameters;
selecting, using the confidence scores, a subset of the unknown data and confident labels, the confident labels comprising ones of the predicted labels for the subset of the unknown data, wherein selecting further comprises:
selecting, as the confident labels, ones of the predicted labels having predicted label values that exceed a threshold value, and
selecting, as the subset of the unknown data, portions of the unknown data corresponding to the confident labels,
wherein the subset of unknown data is less than the unknown data; and
training, using the accepted data, the subset of the unknown data, and the confident labels, the trained prediction model to generate an updated supervised prediction model, wherein the subset of the unknown data, the confident labels, and the accepted data together comprise augmented data that, when used to train the trained prediction model, improves an accuracy of predictions made by the updated supervised prediction model;
executing the updated supervised prediction model on remaining data to generate updated prediction labels, wherein the remaining data comprises the unknown data less the subset of the unknown data;
executing the confidence model on the remaining data to output other confidence scores for the updated predicted labels;
selecting other confident labels for a second subset of the remaining data, outside the subset of data, wherein:
the other confident labels comprise ones of the updated predicted labels for the second subset for which the other confident scores exceed a second threshold, and
the other confident labels and the second subset of the remaining data together increase the size of the initially available training data to increased training data;
training, using the second subset and the other confident labels, the updated supervised prediction model to generate the deployed prediction model, wherein the deployed prediction model is more accurate than the trained prediction model trained using the initially available training data due to the increased training data; and
deploying the deployed prediction model to an enterprise environment.
9 . A method comprising:
receiving live data comprising unknown labels;
generating an updated supervised prediction model by:
training, using accepted data having known labels, an untrained supervised prediction model to generate a trained prediction model, wherein:
the accepted data comprises less than a predetermined amount of training data to achieve a desired accuracy of the trained prediction model, the accepted data comprises initially available training data, and
the untrained supervised prediction model comprises first parameters;
generating, using the trained prediction model, predicted labels for unknown data;
generating, using a confidence model, confidence scores for the predicted labels, wherein the confidence model comprises a logistic regression algorithm comprising second parameters different than the first parameters;
selecting, using the confidence scores, a subset of the unknown data and confident labels, the confident labels comprising ones of the predicted labels for the subset of the unknown data, wherein selecting further comprises:
selecting, as the confident labels, ones of the predicted labels having predicted label values that exceed a threshold value, and
selecting, as the subset of the unknown data, portions of the unknown data corresponding to the confident labels,
wherein the subset of unknown data is less than the unknown data;
training, using the accepted data, the subset of the unknown data, and the confident labels, the trained prediction model to generate the updated supervised prediction model, wherein the subset of the unknown data, the confident labels, and the accepted data together comprise augmented data that, when used to train the trained prediction model, improves an accuracy of predictions made by the updated supervised prediction model;
generating a deployed supervised prediction model by increasing a size of initially available training data, wherein generating the deployed supervised prediction model further comprises:
executing the updated supervised prediction model on remaining data to generate updated prediction labels, wherein the remaining data comprises the unknown data less the subset of the unknown data;
executing the confidence model on the remaining data to output other confidence scores for the updated predicted labels;
selecting other confident labels for a second subset of the remaining data, outside the subset of data, wherein:
the other confident labels comprise ones of the updated predicted labels for the second subset for which the other confident scores exceed a second threshold, and
the other confident labels and the second subset of the remaining data together increase the size of the initially available training data to increased training data;
training, using the second subset and the other confident labels, the updated supervised prediction model to generate a deployed prediction model, wherein the deployed prediction model is more accurate than the trained prediction model trained using the initially available training data due to the increased training data;
deploying the deployed prediction model to an enterprise environment;
generating a result by inputting the live data to the deployed prediction model; and
presenting the result to a user device.
10 . The method of claim 9 wherein the live data comprises a credit application received from the user device, and wherein presenting the result comprises:
determining a credit value from the predicted labels;
granting or rejecting, automatically, the credit application when the credit value exceeds a credit threshold; and
presenting an indication of grant or reject of the credit application to the user device.