Back to results
US 12,608,648 B1Granted
Systems and methods for using synthetic data to ensure training data is properly representative
AssigneeWells Fargo Bank, N.A.
InventorsVinothkumar Venkataraman, Yang Angelina Yang, Ashutosh Verma, Naveen Gururaja Yeri
Life of this patent
Estimating the term…
Abstract
Systems, apparatuses, methods, and computer program products are disclosed for generating representative training data. An example method includes comparing, by data analysis circuitry, a labeled dataset to a target dataset. The example method also includes generating, by training data circuitry, a training dataset based on the comparison, wherein the training dataset comprises at least a portion of the labeled dataset supplemented by a synthetic dataset. The example method also includes training, by modeling circuitry, a model using the training dataset.
Claims20
A method for training a model using representative training data, the method comprising:
- comparing, by data analysis circuitry, each data point of a labeled dataset to each data point of a target dataset, wherein the labeled dataset comprises a plurality of authentic data points, and wherein the target dataset comprises a plurality of data points representing characteristics of a target population for which the model is to generate predictions;
- determining, by the data analysis circuitry, a respective similarity score for each comparison, including a first similarity score for a first comparison of a first data point of the labeled dataset and a first data point of the target dataset;
- generating, by training data circuitry, a training dataset that is fully representative of the target dataset and comprises at least a portion of the plurality of authentic data points of the labeled dataset supplemented by a synthetic dataset, wherein generating the training dataset comprises:
- performing, by the training data circuitry, a data point extraction process for each comparison, including at least:
- determining, by the training data circuitry, that the first similarity score satisfies a predefined threshold, and in response to determining that the first similarity score satisfies the predefined threshold:
- populating, by the training data circuitry, the training dataset with the first data point of the labeled dataset, and
- removing, by the training data circuitry, the first data point of the target dataset from the target dataset;
- identifying, by the data analysis circuitry, one or more data points remaining in the target dataset after the data point extraction process is performed;
- generating, by synthetic data generation circuitry, the synthetic dataset, wherein the synthetic dataset comprises one or more synthetic data points representative of the one or more data points remaining in the target dataset;
- and
- populating, by the training data circuitry, the training dataset with the synthetic dataset;
- and
- training, by modeling circuitry, the model using the training dataset.
10 more elements
An apparatus for training a model using representative training data, the apparatus comprising:
- data analysis circuitry configured to:
- compare each data point of a labeled dataset to each data point of a target dataset, wherein the labeled dataset comprises a plurality of authentic data points, and wherein the target dataset comprises a plurality of data points representing characteristics of a target population for which the model is to generate predictions, and
- determine a respective similarity score for each comparison, including a first similarity score for a first comparison of a first data point of the labeled dataset and a first data point of the target dataset;
- training data circuitry configured to:
- generate a training dataset that is fully representative of the target dataset and comprises at least a portion of the plurality of authentic data points of the labeled dataset supplemented by a synthetic dataset by:
- performing a data point extraction process for each comparison, including at least:
- determining that the first similarity score satisfies a predefined threshold, and in response to determining that the first similarity score satisfies the predefined threshold:
- populating the training dataset with the first data point of the labeled dataset, and
- removing the first data point of the target dataset from the target dataset;
- wherein the data analysis circuitry is further configured to identify one or more data points remaining in the target dataset after the data point extraction process is performed, and
- wherein the apparatus further comprises synthetic data generation circuitry configured to generate the synthetic dataset, wherein the synthetic dataset comprises one or more synthetic data points representative of the one or more data points remaining in the target dataset;
- wherein the training data circuitry is further configured to populate the training dataset with the synthetic dataset;
- and
- wherein the apparatus further comprises modeling circuitry configured to train the model using the training dataset.
11 more elements
A computer program product for training a model using representative training data, the computer program product comprising at least one non-transitory computer-readable storage medium storing software instructions that, when executed, cause an apparatus to:
- compare each data point of a labeled dataset to each data point of a target dataset, wherein the labeled dataset comprises a plurality of authentic data points, and wherein the target dataset comprises a plurality of data points representing characteristics of a target population for which the model is to generate predictions;
- determine a respective similarity score for each comparison, including a first similarity score for a first comparison of a first data point of the labeled dataset and a first data point of the target dataset;
- generate a training dataset that is fully representative of the target dataset and comprises at least a portion of the plurality of authentic data points of the labeled dataset supplemented by a synthetic dataset by:
- performing a data point extraction process for each comparison, including at least:
- determining that the first similarity score satisfies a predefined threshold, and in response to determining that the first similarity score satisfies the predefined threshold:
- populating the training dataset with the first data point of the labeled dataset, and
- removing the first data point of the target dataset from the target dataset;
- identifying one or more data points remaining in the target dataset after the data point extraction process is performed;
- generating, by synthetic data generation circuitry, the synthetic dataset, wherein the synthetic dataset comprises one or more synthetic data points representative of the one or more data points remaining in the target dataset;
- and
- populating the training dataset with the synthetic dataset;
- and
- train the model using the training dataset.
10 more elements
Dependencies are inferred from the claim text.