Memorization model for context violations

1. A computer-implemented method performed by an online system, the method comprising:

generating a memorization model for predicting likelihood scores of context violations for content campaigns, the generating comprising:

identifying a plurality of item feature vectors for a plurality of content campaigns with known context violations, each item feature vector determined from a respective category of content campaign characteristics from a plurality of categories;

training an item weight for each of the plurality of item feature vectors of each category, the item weight indicating a relevance of each item feature vector when it matches a feature vector of a content campaign; and

training a category weight for each category, the category weight indicating a relevance of each category when an item feature vector determined from the category matches a feature vector of a content campaign; and

receiving a content campaign for content to be presented to users of the online system;

identifying a plurality of feature vectors for the received content campaign, each feature vector of the plurality of feature vectors determined from a respective category from the plurality of categories;

predicting a likelihood score of context violations of the content campaign using the memorization model based on the plurality of feature vectors, the predicting comprising:

for each feature vector of the plurality of feature vectors:

comparing the feature vector to the plurality of item feature vectors identified by the memorization model;

identifying a matching item feature of the respective category of the feature vector based on the comparison; and

generating a combination of a trained category weight of the respective category of the feature vector and a trained item weight of the matching item feature vector;

determining the predicted likelihood score based on the combinations determined for the plurality of feature vectors;

determining the received content campaign violates a context policy of the online system based on the predicted likelihood score being included in a likelihood score range; and

performing an action corresponding to the likelihood score range, the action addressing the violation of the context policy by the content campaign.

2. The computer-implemented method of claim 1 , wherein each feature vector of the plurality of feature vectors comprises:

characteristics of content of the content campaign determined from the respective category of content campaign characteristics; and

characteristics of a context of the content campaign determined from the respective category of content campaign characteristics.

3. The computer-implemented method of claim 1 , wherein the predicted likelihood score is a linear combination of the combinations determined for the plurality of feature vectors.

4. The computer-implemented method of claim 1 , wherein the action is selected from a group consisting of: disabling or deleting the content campaign, disabling or deleting information associated with the content campaign, ignoring the content campaign, reviewing the content campaign, and any combination thereof.

5. The computer-implemented method of claim 1 , wherein performing the action comprises:

comparing the predicted likelihood score to a plurality of likelihood score ranges including the likelihood score range, each likelihood score range associated with an action; and

selecting the action corresponding to the likelihood score range including the predicted likelihood score from the plurality of likelihood score ranges.

6. The computer-implemented method of claim 5 , wherein performing the action further comprises:

comparing an additional predicted likelihood score of context violations of the content campaign with the plurality of likelihood score ranges, wherein the additional predicted likelihood score is generated by an additional prediction model; and

selecting an action corresponding to a combination of the likelihood score range that includes the predicted likelihood score and a likelihood score range that includes the additional predicted likelihood score.

7. The computer-implemented method of claim 1 , wherein generating the memorization model further comprises:

receiving the plurality of content campaigns with known context violations;

identifying a set of item feature vectors from each of the plurality of content campaigns;

generating a training set based on the sets of item feature vectors, the training set comprising a positive set including item feature vectors identified for content campaigns of the plurality of content campaigns which violate account terms of the online system; and

training, based on the training set, each item weight associated with each of the plurality of item feature vectors and each category weight associated with each of the plurality of categories.

8. The computer-implemented method of claim 7 , wherein the item weights for each of the plurality of item feature vectors and the category weights for each of the plurality of categories are trained jointly.

9. The computer-implemented method of claim 7 , wherein each of the plurality of content campaigns are labeled based on a likelihood score of context violations of each content campaign determined by an additional prediction model.

10. A non-transitory computer-readable medium comprising computer program instructions that when executed by a computer processor of an online system causes the processor to perform steps comprising:

generating a memorization model for predicting likelihood scores of context violations for content campaigns, the generating comprising:

identifying a plurality of item feature vectors for a plurality of content campaigns with known context violations, each item feature vector determined from a respective category of content campaign characteristics from a plurality of categories;

training an item weight for each of the plurality of item feature vectors of each category, the item weight indicating a relevance of each item feature vector when it matches a feature vector of a content campaign; and

training a category weight for each category, the category weight indicating a relevance of each category when an item feature vector determined from the category matches a feature vector of a content campaign; and

receiving a content campaign for content to be presented to users of the online system;

identifying a plurality of feature vectors for the received content campaign, each feature vector of the plurality of feature vectors determined from a respective category from the plurality of categories;

predicting a likelihood score of context violations of the content campaign using the memorization model based on the plurality of feature vectors, the predicting comprising:

for each feature vector of the plurality of feature vectors:

comparing the feature vector to the plurality of item feature vectors identified by the memorization model;

identifying a matching item feature of the respective category of the feature vector based on the comparison; and

generating a combination of a trained category weight of the respective category of the feature vector and a trained item weight of the matching item feature vector;

determining the predicted likelihood score based on the combinations determined for the plurality of feature vectors;

determining the received content campaign violates a context policy of the online system based on the predicted likelihood score being included in a likelihood score range; and

performing an action corresponding to the likelihood score range, the action addressing the violation of the context policy by the content campaign.

11. The non-transitory computer-readable medium of claim 10 , wherein each feature vector of the plurality of feature vectors comprises:

characteristics of content of the content campaign determined from the respective category of content campaign characteristics; and

characteristics of a context of the content campaign determined from the respective category of content campaign characteristics.

12. The non-transitory computer-readable medium of claim 10 , wherein the predicted likelihood score is a linear combination of the combinations determined for the plurality of feature vectors.

13. The non-transitory computer-readable medium of claim 10 , wherein the action is selected from a group consisting of: disabling or deleting the content campaign, disabling or deleting information associated with the content campaign, ignoring the content campaign, reviewing the content campaign, and any combination thereof.

14. The non-transitory computer-readable medium of claim 10 , wherein performing the action comprises:

comparing the predicted likelihood score to a plurality of likelihood score ranges including the likelihood score range, each likelihood score range associated with an action; and

selecting the action corresponding to the likelihood score range including the predicted likelihood score from the plurality of likelihood score ranges.

15. The non-transitory computer-readable medium of claim 14 , wherein performing the action further comprises:

comparing an additional predicted likelihood score of context violations of the content campaign with the plurality of likelihood score ranges, wherein the additional predicted likelihood score is generated by an additional prediction model; and

selecting an action corresponding to a combination of the likelihood score range that includes the predicted likelihood score and a likelihood score range that includes the additional predicted likelihood score.

16. The non-transitory computer-readable medium of claim 10 , wherein generating the memorization model further comprises:

receiving the plurality of content campaigns with known context violations;

identifying a set of item feature vectors from each of the plurality of content campaigns;

generating a training set based on the sets of item feature vectors, the training set comprising a positive set including item feature vectors identified for content campaigns of the plurality of content campaigns which violate account terms of the online system; and

training, based on the training set, each item weight associated with each of the plurality of item feature vectors and each category weight associated with each of the plurality of categories.

17. The non-transitory computer-readable medium of claim 16 , wherein the item weights for each of the plurality of item feature vectors and the category weights for each of the plurality of categories are trained jointly.

18. The non-transitory computer-readable medium of claim 16 , wherein each of the plurality of content campaigns are labeled based on a likelihood score of context violations of each content campaign determined by an additional prediction model.