Dynamic data selection for a machine learning model
Embodiments implement a machine learning prediction model with dynamic data selection. A number of data predictions generated by a trained machine learning model can be accessed, where the data predictions include corresponding observed data. An accuracy for the machine learning model can be calculated based on the accessed number of data predictions and the corresponding observed data. The accessing and calculating can be iterated using a variable number of data predictions, where the variable number of data predictions is adjusted based on an action taken during a previous iteration, and, when the calculated accuracy fails to meet an accuracy criteria during a given iteration, a training for the machine learning model can be triggered.
1 . A method for implementing a machine learning prediction model with dynamic data selection, the method comprising:
accessing a number of data predictions comprising predictions of data points, the data predictions being generated by a trained machine learning model deployed at a deployment time, wherein actual data values that correspond to the data points of the data predictions are stored;
calculating an accuracy for the machine learning model based on the accessed number of data predictions and the corresponding actual data values;
iterating the accessing and calculating using a variable number of data predictions, wherein,
a defined algorithm automatically adjusts the variable number of data predictions that are used to calculate the accuracy of the machine learning model between successive iterations, the variable number of data predictions between successive iterations being adjusted by a delta value that is calculated by applying a factor value;
when the iterating begins the factor value is a numeric value greater than one, and the factor value per iteration is systematically decreased number of iterations increases relative to the deployment time;
for a given iteration of the iterations, the defined algorithm applies a factor value for the given iteration using a first application type to calculate the delta value when training was triggered for an iteration prior to the given iteration, and the defined algorithm applies the factor value for the given iteration using a second application type to calculate the delta value when training was not triggered for the iteration prior to the given iteration; and
when the calculated accuracy during the given iteration fails to meet an accuracy criteria, a training for the machine learning model is triggered;
terminating, in response to the factor value meeting a criteria value, the iterating of the accessing and calculating such that a configured number of data predictions is determined based on the iterations, wherein training of the machine learning model is triggered by multiple of the iterations; and
generating, using the machine learning model after the terminating, additional predictions, wherein the configured number of data predictions is used to calculate an accuracy for the additional predictions and the calculated accuracy of the additional predictions is used to trigger training of the machine learning model.
2 . The method of claim 1 , wherein the iteration prior to the given iteration comprises an iteration directly preceding the given iteration.
3 . The method of claim 1 , wherein the triggered training comprises a retraining or updated training for the trained machine learning model.
4 . The method of claim 1 , wherein, when training is triggered, next iterations of the accessing and calculating use data predictions generated by the machine learning model generated trained via by the triggered training.
5 . The method of claim 1 , wherein the iterating is performed according to a predetermined period, and the predetermined period is a predetermined period of time or a predetermined amount of data predictions that have corresponding actual data values.
6 . The method of claim 1 , wherein the variable number of data predictions for the given iteration is increased when training was triggered during the iteration prior to the given iteration and decreased when training was not triggered during the iteration prior to the given iteration.
7 . The method of claim 1 , wherein,
the variable number of data predictions for the given iteration is based on the variable number of data predictions for the iteration prior to the given iteration;
the variable number of data predictions for the given iteration comprises the variable number of data predictions for the iteration prior to the given iteration multiplied by the factor value for the given iteration when training was triggered during the iteration prior to the given iteration; and
the variable number of data predictions for the given iteration comprises the variable number of data predictions for the iteration prior to the given iteration divided by the factor value for the given iteration when a training was not triggered during the iteration prior to the given iteration.
8 . A system for implementing a machine learning prediction model with dynamic data selection, the system comprising:
a processor; and
a memory storing instructions for execution by the processor, the instructions configuring the processor to:
access a number of data predictions comprising predictions of data points, the data predictions being generated by a trained machine learning model deployed at a deployment time, wherein actual data values that correspond to the data points of the data predictions are stored;
calculate an accuracy for the machine learning model based on the accessed number of data predictions and the corresponding actual data values;
iterate the accessing and calculating using a variable number of data predictions, wherein,
a defined algorithm automatically adjusts the variable number of data predictions that are used to calculate the accuracy of the machine learning model between successive iterations, the variable number of data predictions between successive iterations being adjusted by a delta value that is calculated by applying a factor value;
when the iterating begins the factor value is a numeric value greater than one, and the factor value per iteration is systematically decreased as a number of iterations increases relative to the deployment time;
for a given iteration of the iterations, the defined algorithm applies a factor value for the given iteration using a first application type to calculate the delta value when training was triggered for an iteration prior to the given iteration, and the defined algorithm applies the factor value for the given iteration using a second application type to calculate the delta value when training was not triggered for the iteration prior to the given iteration; and
when the calculated accuracy during the given iteration fails to meet an accuracy criteria, a training for the machine learning model is triggered;
terminate, in response to the factor value meeting a criteria value, the iterating of the accessing and calculating such that a configured number of data predictions is determined based on the iterations, wherein training of the machine learning model is triggered by multiple of the iterations; and
generate, using the machine learning model after the terminating, additional predictions, wherein the configured number of data predictions is used to calculate an accuracy for the additional predictions and the calculated accuracy of the additional predictions is used to trigger training of the machine learning model.
9 . The system of claim 8 , wherein the iteration prior to the given iteration comprises an iteration directly preceding the given iteration.
10 . The system of claim 9 , wherein the iterating is performed according to a predetermined period, and the predetermined period is a predetermined period of time or a predetermined amount of data predictions that have corresponding actual data values.
11 . The system of claim 10 , wherein the variable number of data predictions for the given iteration is increased when training was triggered during the iteration prior to the given iteration and the variable number of data predictions is decreased when training was not triggered during the iteration prior to the given iteration.
12 . A non-transitory computer readable medium having instructions stored thereon that, when executed by a processor, cause the processor to implement a machine learning prediction model with dynamic data selection, wherein, when executed, the instructions cause the processor to:
access a number of data predictions comprising predictions of data points, the data predictions being generated by a trained machine learning model deployed at a deployment time, wherein the data predictions comprise corresponding actual data values of the data points;
calculate an accuracy for the machine learning model based on the accessed number of data predictions and the corresponding actual data values;
iterate the accessing and calculating using a variable number of data predictions, wherein,
a defined algorithm automatically adjusts the variable number of data predictions that are used to calculate the accuracy of the machine learning model between successive iterations, the variable number of data predictions between successive iterations being adjusted by a delta value that is calculated by applying a factor value;
when the iterating begins the factor value is a numeric value greater than one, and the factor value per iteration is systematically decreased as a number of iterations increases relative to the deployment time; and
for a given iteration of the iterations, the defined algorithm applies a factor value for the given iteration using a first application type to calculate the delta value when training was triggered for an iteration prior to the given iteration, and the defined algorithm applies the factor value for the given iteration using a second application type to calculate the delta value when training was not triggered for the iteration prior to the given iteration,
when the calculated accuracy during the given iteration fails to meet an accuracy criteria, a training for the machine learning model is triggered;
terminate, in response to the factor value meeting a criteria value, the iterating of the accessing and calculating such that a configured number of data predictions is determined based on the iterations, wherein training of the machine learning model is triggered by multiple of the iterations; and
generate, using the machine learning model after the terminating, additional predictions, wherein the configured number of data predictions is used to calculate an accuracy for the additional predictions and the calculated accuracy of the additional predictions is used to trigger training of the machine learning model.
13 . The method of claim 1 , wherein applying the factor value for the given iteration using the first application type calculates a delta value that increases the variable number of data points between the iteration prior to the given iteration and the given iteration, and applying the factor value for the given iteration using the second application type calculates a delta value that decreases the variable number of data points between the iteration prior to the given iteration and the given iteration.
14 . The system of claim 8 , wherein applying the factor value for the given iteration using the first application type calculates a delta value that increases the variable number of data points between the iteration prior to the given iteration and the given iteration, and applying the factor value for the given iteration using the second application type calculates a delta value that decreases the variable number of data points between the iteration prior to the given iteration and the given iteration.
15 . The method of claim 1 , wherein the factor value per iteration is systematically decreased as the number of iterations increases relative to the deployment time by applying a step value to the factor value between successive iterations.
16 . The method of claim 15 , wherein applying the step value to the factor value over the iterations decreases the factor value to the criteria value or below the criteria value by a time when the iterating is terminated.
17 . The method of claim 15 , wherein the step value is a numeric value less than one and applying the step value to the factor value comprises multiplying the factor value by the step value.
18 . The method of claim 1 , wherein the machine learning model comprises a demand forecast model, the generated additional predictions comprise demand forecasts, and product shipments are executed in response to the generated additional predictions.
19 . The system of claim 8 , wherein the factor value per iteration is systematically decreased as the number of iterations increases relative to the deployment time by applying a step value to the factor value between successive iterations.
20 . The system of claim 19 , wherein applying the step value to the factor value over the iterations decreases the factor value to the criteria value or below the criteria value by a time when the iterating is terminated.