Meta-model classifier

1. A method comprising:

accessing, by an online system, a model attribute store, the model attribute store identifying configuration information and model performance scores for a plurality of models, each model used to predict performance metrics regarding content from a third party system presented to users of the online system;

accessing a meta-model classifier, the meta-model classifier trained using the plurality of models in the model attribute store to predict for a candidate model a predicted model performance score of that candidate model, wherein the meta-model classifier is separately trained from the plurality of models;

generating a plurality of candidate models for input to the meta-model classifier, each of the plurality of candidate models including a distinct set of configuration information;

computing a predicted model performance score for a selected candidate model in the plurality of candidate models using the meta-model classifier; and

transmitting a report to the third party system indicating predicted model performance score for the selected candidate model.

2. The method of claim 1 , wherein the model attribute store further includes, for each training model, features used by the training model and corresponding feature importance scores, each feature importance score indicating an impact each corresponding feature has on influencing the predicted performance metric generated by the model, wherein the meta-model classifier is further trained using the features and feature importance scores as input, and wherein each of the generated candidate models includes features used by the candidate model.

3. The method of claim 2 , wherein the features include data extracted from a user profile store of the online system.

4. The method of claim 1 , wherein the model performance score indicates a value of a corresponding model to the third party system.

5. The method of claim 4 , wherein value is a measure of time spent by the user at the third party system.

6. The method of claim 1 , wherein the model performance score indicates an accuracy of the predicted performance metric made by a corresponding model compared to a measured performance metric.

7. The method of claim 1 , wherein the computing the predicted model performance score further comprises:

extracting a set of input features for the meta-model classifier from the candidate model, the input features for the meta-model classifier including the configuration information of the candidate model;

setting as zero those input features for the meta-model predictor that are not present in the extracted set of input features; and

computing the predicted model performance score for the selected candidate model using the extracted set of input features and the zeroed input features.

8. The method of claim 1 , further comprising:

removing the candidate model in response to the predicted model performance score of the candidate model falling below a threshold score.

9. The method of claim 1 , wherein the generating a plurality of candidate models further comprises:

modifying a configuration parameter of the configuration information of a training model in the model attribute store to generate a mutated configuration information; and

generating a new candidate model with the mutated configuration information.

10. The method of claim 1 , wherein the generating a plurality of candidate models further comprises:

adding a configuration parameter of the configuration information of a training model in the model attribute store to generate a new configuration information; and

generating a new candidate model with the new configuration information.

11. A computer program product comprising a non-transitory computer readable storage medium having instructions encoded thereon that, when executed by a processor, cause the processor to:

access, by an online system, a model attribute store, the model attribute store identifying configuration information and model performance scores for a plurality of models, each model used to predict performance metrics regarding content from a third party system presented to users of the online system;

access a meta-model classifier, the meta-model classifier trained using the plurality of models in the model attribute store to predict for a candidate model a predicted model performance score of that candidate model, wherein the meta-model classifier is separately trained from the plurality of models;

generate a plurality of candidate models for input to the meta-model classifier, each of the plurality of candidate models including a distinct set of configuration information;

compute a predicted model performance score for a selected candidate model in the plurality of candidate models using the meta-model classifier; and

transmit a report to the third party system indicating predicted model performance score for the selected candidate model.

12. The computer program product of claim 11 , wherein the model attribute store further includes, for each training model, features used by the training model and corresponding feature importance scores, each feature importance score indicating an impact each corresponding feature has on influencing the predicted performance metric generated by the model, wherein the meta-model classifier is further trained using the features and feature importance scores as input, and wherein each of the generated candidate models includes features used by the candidate model.

13. The computer program product of claim 12 , wherein the features include data extracted from a user profile store of the online system.

14. The computer program product of claim 11 , wherein the model performance score indicates a value of a corresponding model to the third party system.

15. The computer program product of claim 14 , wherein value is a measure of time spent by the user at the third party system.

16. The computer program product of claim 11 , wherein the model performance score indicates an accuracy of the predicted performance metric made by a corresponding model compared to a measured performance metric.

17. The computer program product of claim 11 , wherein the non-transitory computer readable storage medium has further instructions encoded thereon for the computation of the predicted model performance score that, when executed by the processor, cause the processor to:

extract a set of input features for the meta-model classifier from the candidate model, the input features for the meta-model classifier including the configuration information of the candidate model;

set as zero those input features for the meta-model predictor that are not present in the extracted set of input features; and

compute the predicted model performance score for the selected candidate model using the extracted set of input features and the zeroed input features.

18. The computer program product of claim 11 , wherein the non-transitory computer readable storage medium has further instructions encoded thereon that, when executed by the processor, cause the processor to:

remove the candidate model in response to the predicted model performance score of the candidate model falling below a threshold score.

19. The computer program product of claim 11 , wherein the non-transitory computer readable storage medium has further instructions encoded thereon for the generation of a plurality of candidate models that, when executed by the processor, cause the processor to:

modify a configuration parameter of the configuration information of a training model in the model attribute store to generate a mutated configuration information; and

generate a new candidate model with the mutated configuration information.

20. The computer program product of claim 11 , wherein the non-transitory computer readable storage medium has further instructions encoded thereon for the generation of the plurality of candidate models that, when executed by the processor, cause the processor to:

add a configuration parameter of the configuration information of a training model in the model attribute store to generate a new configuration information; and

generate a new candidate model with the new configuration information.