Techniques to use machine learning for risk management
AssigneeCapital One Services, LLC
InventorsYvette Jackson, Jeffrey Capelli, Jennelle Spurlock
Life of this patent
Estimating the term…
Abstract
Various embodiments are generally directed to techniques to build and train a machine learning model with features of which at least one feature corresponds to risk indicia and at least one other feature corresponds to a data source. These features, in general, provide data (e.g., values) indicating a degree of relevance between a particular record and a risk assessment of that particular record's subject matter. User refinement in the form of user selections and other interactions with the particular record and other records provide insights into proper risk management.
Claims20
An apparatus, comprising:
- a processing circuit;
- and
- logic stored in computer memory and executed on the processing circuit, the logic operative to cause the processing circuit to:
- perform feature extraction on records across a plurality of datasets, features corresponding to the feature extraction comprise at least one feature associated with risk indicia and at least one feature associated with a particular data source of at least one of the plurality of datasets;
- build a machine learning model based upon data generated during the feature extraction on the records, the machine learning model comprising risk assessment data for the records across the plurality of datasets;
- identify a set of records from the records across the plurality of datasets, each record of the set of records having risk assessment data that exceeds a baseline threshold;
- return at least one record of the set of records in response to search queries;
- and
- modify the risk assessment data in the machine learning model in response to user selections corresponding to the at least one returned record.
A computer-implemented method executed on a processing circuit, comprising:
- generating feature extraction data from records across a plurality of datasets, features corresponding to the feature extraction data comprise at least one feature associated with risk indicia and at least one feature associated with a particular data source of at least one of the plurality of datasets;
- building a machine learning model based upon the feature extraction data on the records, the machine learning model comprising risk assessment data for the records across the plurality of datasets;
- identifying a set of records from the records across the plurality of datasets, each record of the set of records having risk assessment data that exceeds a baseline threshold;
- returning at least one record of the set of records in response to search queries;
- and
- modifying the risk assessment data in the machine learning model in response to user selections corresponding to the at least one returned record.
At least one non-transitory computer-readable storage medium comprising instructions that, when executed, cause a system to:
- generate feature extraction data from records across a plurality of datasets, features corresponding to the feature extraction data comprise at least one feature associated with risk indicia and at least one feature associated with a particular data source of at least one of the plurality of datasets;
- build a machine learning model based upon the feature extraction data on the records, the machine learning model comprising risk assessment data for the records across the plurality of datasets;
- execute a clustering process on the records across the plurality of datasets to identify a set of records, each record of the set of records having risk assessment data that exceeds a baseline threshold;
- return at least one record of the set of records in response to search queries;
- modify the risk assessment data in the machine learning model in response to user selections corresponding to the at least one returned record;
- and
- execute the clustering process on the records across the plurality of datasets to identify a new set of records.
Dependencies are inferred from the claim text.