Sensitivity in supervised machine learning with experience data
In an example embodiment, a process is introduced into a machine learned model where additional results are output by the machine learned model in addition to those results that would be obtained through use of the trained model itself. In some example embodiments, these additional results may be random or semi-random to introduce results that might otherwise not have been recommended by the machine learned model. By introducing such additional results in a controlled way, it becomes possible to reduce biases caused by a self-reinforcing feedback loop while still presenting users with accurate machine learned model results.
1. A system comprising:
at least one hardware processor; and
a computer-readable medium storing instructions that, when executed by the at least one hardware processor, cause the at least one hardware processor to perform operations comprising:
training a machine learned model using a machine learning algorithm, the training utilizing a first set of training data stored in a database;
receiving, via an application programming interface (API), a request from a component for recommended values of a first type, the request generated based on first user interaction with a user interface;
calculating, using the machine learned model, a set of one or more recommended values of the first type;
identifying a set of additional values of the first type;
receiving feedback about at least one value in the set of additional values, the feedback having been generated based on second user interaction with the user interface;
tagging the at least one value with a feedback event and an indication of a version number of the machine learned model that was in use at a time the feedback was received;
retraining the machine learned model based on the feedback; and
testing the retrained machine learned model against a set of conditions that was in place at the time the feedback was received, using the feedback event and indication of the version number of the machine learned model.
2. The system of claim 1 , wherein the set of additional values include only values randomly selected from a subset of possible values of the first type.
3. The system of claim 2 , wherein the subset of possible values of the first type includes only values containing one or more attribute values matching preconfigured criteria for an attribute of values of the first type.
4. The system of claim 1 , wherein the feedback includes positive interaction with the at least one value in a user interface.
5. The system of claim 1 , wherein the feedback includes an explicit rating provided by a user for each of the at least one value.
6. The system of claim 1 , wherein the feedback includes a key performance indicator (KPI) related to each of the at least one value.
7. The system of claim 1 , wherein the calculating is performed in response to receiving a first call from a user interface and wherein the identifying is performed in response to receiving a second call from the user interface.
8. A method comprising:
training a machine learned model using a machine learning algorithm, the training utilizing a first set of training data stored in a database;
receiving, via an application programming interface (API), a request from a component for recommended values of a first type, the request generated based on first user interaction with a user interface;
calculating, using the machine learned model, a set of one or more recommended values of the first type;
identifying a set of additional values of the first type;
receiving feedback about at least one value in the set of additional values, the feedback having been generated based on second user interaction with the user interface;
tagging the at least one value with a feedback event and an indication of a version number of the machine learned model that was in use at a time the feedback was received;
retraining the machine learned model based on the feedback; and
testing the retrained machine learned model against a set of conditions that was in place at the time the feedback was received, using the feedback event and indication of the version number of the machine learned model.
9. The method of claim 8 , wherein the set of additional values include only values randomly selected from a subset of possible values of the first type.
10. The method of claim 9 , wherein the subset of possible values of the first type includes only values containing one or more attribute values matching preconfigured criteria for an attribute of values of the first type.
11. The method of claim 8 , wherein the feedback includes positive interaction with the at least one value in a user interface.
12. The method of claim 8 , wherein the feedback includes an explicit rating provided by a user for each of the at least one value.
13. The method of claim 8 , wherein the feedback includes a key performance indicator (KPI) related to each of the at least one value.
14. The method of claim 8 , wherein the calculating is performed in response to receiving a first call from a user interface and wherein the identifying is performed in response to receiving a second call from the user interface.
15. A non-transitory machine-readable medium storing instructions which, when executed by one or more processors, cause the one or more processors to perform operations comprising:
training a machine learned model using a machine learning algorithm, the training utilizing a first set of training data stored in a database;
receiving, via an application programming interface (API), a request from a component for recommended values of a first type, the request generated based on first user interaction with a user interface;
calculating, using the machine learned model, a set of one or more recommended values of the first type;
identifying a set of additional values of the first type;
receiving feedback about at least one value in the set of additional values, the feedback having been generated based on second user interaction with the user interface;
tagging the at least one value with a feedback event and an indication of a version number of the machine learned model that was in use at a time the feedback was received;
retraining the machine learned model based on the feedback; and
testing the retrained machine learned model against a set of conditions that was in place at the time the feedback was received, using the feedback event and indication of the version number of the machine learned model.
16. The non-transitory machine-readable medium of claim 15 , wherein the set of additional values include only values randomly selected from a subset of possible values of the first type.
17. The non-transitory machine-readable medium of claim 16 , wherein the subset of possible values of the first type includes only values containing one or more attribute values matching preconfigured criteria for an attribute of values of the first type.
18. The non-transitory machine-readable medium of claim 15 , wherein the feedback includes positive interaction with the at least one value in a user interface.
19. The non-transitory machine-readable medium of claim 15 , wherein the feedback includes an explicit rating provided by a user for each of the at least one value.
20. The non-transitory machine-readable medium of claim 15 , wherein the feedback includes a key performance indicator (KPI) related to each of the at least one value.