Template regularization for generalization of learning systems
View Patent ↗Systems and techniques are disclosed for training a machine learning model based on one or more regularization penalties associated with one or more features. A template having a lower regularization penalty may be given preference over a template having a higher regularization penalty. A regularization penalty may be determined based on domain knowledge. A restrictive regularization penalty may be assigned to a template based on determining that a template occurrence is below a stability threshold and may be modified if the template occurrence meets or exceeds the stability threshold.
1. A computer-implemented method of training a machine learning model on labeled examples, wherein the machine learning model is configured to receive an example having a plurality of features and to generate a predicted output for the received example, the method comprising:
obtaining data defining a plurality of templates, wherein each template corresponds to one or more categories of features;
assigning a respective regularization penalty to each of the plurality of templates; and
training the machine learning model on the labeled examples, comprising, for each labeled example and for each of the plurality of templates:
determining, using the machine learning model, a respective weight for the template based on the features of the labeled example that belong to the one or more categories that correspond to the template, and
modifying the respective weight for the template by applying the respective regularization penalty for the template to the respective weight for the template determined by the machine learning model,
wherein, during the training, a template having a lower regularization penalty is emphasized over a template having a higher regularization penalty.
2. The method of claim 1 , wherein the respective regularization penalty for each of the templates is based on domain knowledge.
3. The method of claim 2 , wherein the domain knowledge corresponds to historic data associated with at least one feature associated with the template.
4. The method of claim 2 , wherein the domain knowledge is provided by a user.
5. The method of claim 1 , further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template is below a stability threshold; and
assigning a restrictive regularization penalty to the first template based on the determination.
6. The method of claim 1 , further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template meets or exceeds a stability threshold; and
modifying the regularization penalty for the first template from a higher regularization penalty to a lower regularization penalty, based on the determination.
7. The method of claim 1 , wherein the example characterizes a setting for presenting a content item to a user and the predicted output is a prediction of a likelihood of a user selection of the content item.
8. The method of claim 7 , further comprising selecting a content item to provide for presentation to the user based on the predicted output.
9. A system comprising:
one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:
obtaining data defining a plurality of templates, wherein each template corresponds to one or more categories of features;
assigning a respective regularization penalty to each of the plurality of templates; and
training the machine learning model on the labeled examples, comprising, for each labeled example and for each of the plurality of templates:
determining, using the machine learning model, a respective weight for the template based on the features of the labeled example that belong to the one or more categories that correspond to the template, and
modifying the respective weight for the template by applying the respective regularization penalty for the template to the respective weight for the template determined by the machine learning model,
wherein, during the training, a template having a lower regularization penalty is emphasized over a template having a higher regularization penalty.
10. The system of claim 9 , wherein the respective regularization penalty for each of the templates is based on domain knowledge.
11. The system of claim 10 , wherein the domain knowledge corresponds to historic data associated with at least one feature associated with the template.
12. The system of claim 10 , wherein the domain knowledge is provided by a user.
13. The system of claim 9 , the operations further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template is below a stability threshold; and
assigning a restrictive regularization penalty to the first template based on the determination.
14. The system of claim 9 , the operations further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template meets or exceeds a stability threshold; and
modifying the regularization penalty for the first template from a higher regularization penalty to a lower regularization penalty, based on the determination.
15. The system of claim 9 , wherein the example characterizes a setting for presenting a content item to a user and the predicted output is a prediction of a likelihood of a user selection of the content item.
16. The system of claim 15 , the operations further comprising selecting a content item to provide for presentation to the user based on predicted output.
17. A non-transitory computer readable medium encoded with a computer program comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
obtaining data defining a plurality of templates, wherein each template corresponds to one or more categories of features;
assigning a respective regularization penalty to each of the plurality of templates; and
training the machine learning model on the labeled examples, comprising, for each labeled example and for each of the plurality of templates:
determining, using the machine learning model, a respective weight for the template based on the features of the labeled example that belong to the one or more categories that correspond to the template, and
modifying the respective weight for the template by applying the respective regularization penalty for the template to the respective weight for the template determined by the machine learning model,
wherein, during the training, a template having a lower regularization penalty is emphasized over a template having a higher regularization penalty.
18. The non-transitory computer readable medium of claim 17 , wherein the respective regularization penalty for each of the templates is based on domain knowledge.
19. The non-transitory computer readable medium of claim 17 , the operations further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template is below a stability threshold; and
assigning a restrictive regularization penalty to the first template based on the determination.
20. The non-transitory computer readable medium of claim 17 , the operations further comprising:
determining that, for a first template, a number of occurrences of distinct features belonging to the one or more categories corresponding to the first template meets or exceeds a stability threshold; and
modifying the regularization penalty for the first template from a higher regularization penalty to a lower regularization penalty, based on the determination.