Superposition-based content serving
Embodiments of the disclosed technologies receive a request including a user identifier and metadata associated with a slot available at a user system, remove the user identifier from the request to produce anonymized request data, receive, from a machine learning model, superposition data that correlates with the anonymized request data, send the superposition data for the anonymized request data to a real-time content-to-request matching process, receive, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected based on the superposition data, and initiate the selected content distribution through the network to the slot in response to the request.
1. A method comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and
initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process comprises determining that a performance metric associated with the selected content distribution does not satisfy a threshold and applying the selected content distribution to a group label that has a lower confidence score.
2. The method of claim 1 , wherein the real-time content-to-request matching process further comprises:
generating an index that maps the superposition data to targeting criteria;
each targeting criterion is associated with at least one content distribution of the plurality of available content distributions; and
using the index to select the content distribution.
3. The method of claim 2 , wherein generating the index comprises:
using the plurality of different group labels to match the plurality of confidence scores with the targeting criteria.
4. The method of claim 1 , wherein the real-time content-to-request matching process further comprises:
determining a confidence score associated with a group label of the plurality of different group labels that maps to a targeting criterion of a content distribution of the plurality of available content distributions;
determining that the confidence score satisfies a pacing threshold associated with the content distribution; and
selecting the content distribution as the selected content distribution.
5. The method of claim 1 , wherein the real-time content-to-request matching process further comprises:
generating a value of an interaction prediction for a group label of the plurality of different group labels;
adjusting the value of the interaction prediction based on a confidence score associated with the group label; and
selecting the selected content distribution based on the adjusted value of the interaction prediction.
6. The method of claim 1 , wherein the real-time content-to-request matching process comprises:
adjusting values of a serving criterion associated with the plurality of available content distributions based on the superposition data; and
based on the adjusted values of the serving criterion, selecting the selected content distribution.
7. The method of claim 1 , wherein the real-time content-to-request matching process further comprises:
selecting, as the selected content distribution, a content distribution of the plurality of available content distributions that maps to a group label that does not have a highest confidence score of the plurality of confidence scores.
8. The method of claim 1 , wherein the plurality of different group labels comprises a plurality of different job titles.
9. A method comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and
initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process further comprises:
determining a confidence score associated with a group label of the plurality of different group labels that maps to a targeting criterion of a content distribution of the plurality of available content distributions;
determining that the confidence score does not satisfy a pacing threshold associated with the content distribution; and
discontinuing the real-time content-to-request matching process for the content distribution.
10. A method comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and
initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process further comprises:
determining that a first confidence score associated with a group label of the plurality of different group labels is lower than second confidence scores associated with the other group labels of the plurality of different group labels;
based on the first confidence score, adjusting a value of a serving criterion associated with a content distribution that has a targeting criterion that maps to the group label;
determining that the adjusted value of the serving criterion meets or exceeds a threshold serving criterion; and
selecting the content distribution as the selected content distribution.
11. A system comprising:
at least one memory; and
at least one processor coupled to the at least one memory;
wherein the at least one memory comprises instructions that, when executed by the at least one processor cause the at least one processor to perform operations comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process comprises determining that a performance metric associated with the selected content distribution does not satisfy a threshold and applying the selected content distribution to a group label that has a lower confidence score.
12. The system of claim 11 , wherein the real-time content-to-request matching process further comprises:
generating an index that maps the superposition data to targeting criteria;
each targeting criterion is associated with at least one content distribution of the plurality of available content distributions; and
using the index to select the content distribution.
13. The system of claim 11 , wherein the real-time content-to-request matching process further comprises:
determining a first confidence score associated with a group label of the plurality of different group labels that maps to a targeting criterion of a content distribution of the plurality of available content distributions;
determining that the first confidence score satisfies a pacing threshold associated with the content distribution; and
selecting the content distribution as the selected content distribution.
14. The system of claim 11 , wherein the real-time content-to-request matching process further comprises:
generating a value of an interaction prediction for a group label of the plurality of different group labels;
adjusting the value of the interaction prediction based on the confidence score associated with the group label; and
selecting the selected content distribution based on the adjusted value of the interaction prediction.
15. The system of claim 11 , wherein the real-time content-to-request matching process comprises:
adjusting values of a serving criterion associated with the plurality of available content distributions based on the superposition data; and
based on the adjusted values of the serving criterion, selecting the selected content distribution.
16. The system of claim 11 , wherein the real-time content-to-request matching process further comprises:
selecting, as the selected content distribution, a content distribution of the plurality of available content distributions that maps to a group label that does not have a highest confidence score of the plurality of confidence scores.
17. A system comprising:
at least one memory; and
at least one processor coupled to the at least one memory;
wherein the at least one memory comprises instructions that, when executed by the at least one processor cause the at least one processor to perform operations comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and
initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process further comprises:
determining a first confidence score associated with a group label of the plurality of different group labels that maps to a targeting criterion of a content distribution of the plurality of available content distributions;
determining that the first confidence score does not satisfy a pacing threshold associated with the content distribution; and
discontinuing the real-time content-to-request matching process for the content distribution.
18. A system comprising:
at least one memory; and
at least one processor coupled to the at least one memory;
wherein the at least one memory comprises instructions that, when executed by the at least one processor cause the at least one processor to perform operations comprising:
receiving, from a network, a request including a user identifier and metadata associated with a slot available at a user system;
removing the user identifier from the request to produce anonymized request data;
sending the anonymized request data to a machine learning model trained to determine relationships between labels and requests;
receiving, from the machine learning model, superposition data that correlates with the anonymized request data including the metadata;
the superposition data comprises a plurality of different group labels and a plurality of confidence scores, each confidence score is associated with one group label of the plurality of different group labels, and each confidence score is computed independently of the confidence scores computed for other group labels of the plurality of different group labels;
sending the superposition data for the anonymized request data to a real-time content-to-request matching process;
receiving, from the real-time content-to-request matching process, an identifier that identifies a content distribution selected from a plurality of available content distributions based on the superposition data; and
initiating the selected content distribution through the network to the slot in response to the request, wherein the real-time content-to-request matching process further comprises:
determining that a first confidence score associated with a group label of the plurality of different group labels is lower than second confidence scores associated with the other group labels of the plurality of different group labels;
based on the first confidence score, adjusting a value of a serving criterion associated with a content distribution that has a targeting criterion that maps to the group label;
determining that the adjusted value of the serving criterion meets or exceeds a threshold serving criterion; and
selecting the content distribution as the selected content distribution.