Decentralized cross-node learning for audience propensity prediction
Embodiments of the disclosed technologies receive a first-party trained model and a first-party data set from a first-party system into a protected environment, receive a first third-party data set into the protected environment, and, in a data clean room, joining the first-party data set and the first third-party data set to create a joint data set for the particular segment, tuning a first-party trained model with the joint data set to create a third-party tuned model, sending model parameter data learned in the data clean room as a result of the tuning to an aggregator node, receiving a globally tuned version of the first-party trained model from the aggregator node, applying the globally tuned version of the first-party trained model to a second third-party data set to produce a scored third-party data set, and providing the scored third-party data set to a content distribution service of the first-party system.
1 . A method comprising:
receiving a first-party trained model and a first-party data set from a first-party system into a protected environment of the first-party system;
the first-party trained model models propensity correlations between first-party entities and third-party segments;
the protected environment comprises a plurality of third-party nodes and an aggregator node;
a third-party node of the plurality of third-party nodes comprises a data clean room;
the third-party node is associated with a source of a content distribution;
the data clean room is associated with a particular segment;
receiving a first third-party data set for the particular segment into the protected environment; and
in the data clean room of the third-party node, (i) semantically aligning and joining the first-party data set and the first third-party data set to create a joint data set for the particular segment, (ii) tuning the first-party trained model with the joint data set to create a third-party tuned model for the particular segment, (iii) sending model parameter data learned in the data clean room as a result of the tuning to the aggregator node, (iv) receiving a globally tuned version of the first-party trained model from the aggregator node, (v) applying the globally tuned version of the first-party trained model to a second third-party data set for the particular segment to produce a scored third-party data set, and (vi) providing the scored third-party data set to a content distribution service of the first-party system.
2 . The method of claim 1 , further comprising, at the protected environment, mapping the second third-party data set to the third-party node.
3 . The method of claim 1 , further comprising, in the data clean room, using the globally tuned version of the first-party trained model to create the third-party tuned model.
4 . The method of claim 1 , further comprising, at the third-party node, creating a second data clean room in the protected environment for a second particular segment.
5 . The method of claim 1 , further comprising, at the protected environment, creating a second third-party node for a second source of a second content distribution.
6 . The method of claim 5 , further comprising, at the protected environment:
logically isolating the third-party node from the second third-party node.
7 . The method of claim 1 , further comprising, at the protected environment:
logically isolating the aggregator node from the third-party node.
8 . The method of claim 1 , further comprising:
checking an opt-in flag of the third-party node;
when the opt-in flag indicates that the source has opted in to aggregation, sending the model parameter data to the aggregator node; and
when the opt-in flag indicates that the source has not opted in to the aggregation, skipping the sending of the model parameter data to the aggregator node.
9 . The method of claim 1 , further comprising:
checking an opt-in flag of the third-party node;
when the opt-in flag indicates that the source has opted in to aggregation, receiving the globally tuned version of the first-party trained model from the aggregator node; and
when the opt-in flag indicates that the source has not opted in to the aggregation, skipping the receiving of the globally tuned version of the first-party trained model from the aggregator node.
10 . A system comprising:
a processor; and
memory coupled to the processor, wherein the memory comprises instructions that when executed by the processor cause the processor to:
receive a first-party trained model and a first-party data set from a first-party system into a protected environment of the first-party system;
the first-party trained model models propensity correlations between first-party entities and third-party segments;
the protected environment comprises a plurality of third-party nodes and an aggregator node;
a third-party node of the plurality of third-party nodes comprises a data clean room;
the third-party node is associated with a source of a content distribution;
the data clean room is associated with a particular segment;
receive a first third-party data set for the particular segment into the protected environment; and
in the data clean room of the third-party node, (i) semantically align and join the first-party data set and the first third-party data set to create a joint data set for the particular segment, (ii) tune the first-party trained model with the joint data set to create a third-party tuned model for the particular segment, (iii) send model parameter data learned in the data clean room as a result of the tuning to the aggregator node, (iv) receive a globally tuned version of the first-party trained model from the aggregator node, (v) apply the globally tuned version of the first-party trained model to a second third-party data set for the particular segment to produce a scored third-party data set, and (vi) provide the scored third-party data set to a content distribution service of the first-party system.
11 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, at the protected environment, map the second third-party data set to the third-party node.
12 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, in the data clean room, use the globally tuned version of the first-party trained model to create the third-party tuned model.
13 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, at the third-party node, create a second data clean room in the protected environment for a second particular segment.
14 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, at the protected environment, create a second third-party node for a second source of a second content distribution.
15 . The system of claim 14 , wherein the instructions, when executed by the processor, cause the processor to, at the protected environment:
logically isolate the third-party node from the second third-party node.
16 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, at the protected environment:
logically isolate the aggregator node from the third-party node.
17 . The system of claim 10 , wherein the instructions, when executed by the processor, cause the processor to, at the protected environment, at least one of:
map the second third-party data set to the third-party node;
create a second third-party node for a second source of a second content distribution;
logically isolate the third-party node from the second third-party node; or
logically isolate the aggregator node from the third-party node.
18 . A non-transitory computer-readable medium comprising instructions that when executed by a processor cause the processor to:
receive a first-party trained model and a first-party data set from a first-party system into a protected environment of the first-party system;
the first-party trained model models propensity correlations between first-party entities and third-party segments;
the protected environment comprises a plurality of third-party nodes and an aggregator node;
a third-party node of the plurality of third-party nodes comprises a data clean room;
the third-party node is associated with a source of a content distribution;
the data clean room is associated with a particular segment;
receive a first third-party data set for the particular segment into the protected environment; and
in the data clean room of the third-party node, (i) semantically align and join the first-party data set and the first third-party data set to create a joint data set for the particular segment, (ii) tune the first-party trained model with the joint data set to create a third-party tuned model for the particular segment, (iii) send model parameter data learned in the data clean room as a result of the tuning to the aggregator node, (iv) receive a globally tuned version of the first-party trained model from the aggregator node, (v) apply the globally tuned version of the first-party trained model to a second third-party data set for the particular segment to produce a scored third-party data set, and (vi) provide the scored third-party data set to a content distribution service of the first-party system.
19 . The non-transitory computer-readable medium of claim 18 , wherein the instructions, when executed by the processor, cause the processor to:
check an opt-in flag of the third-party node;
when the opt-in flag indicates that the source has opted in to aggregation, send the model parameter data to the aggregator node; and
when the opt-in flag indicates that the source has not opted in to the aggregation, skip the sending of the model parameter data to the aggregator node.
20 . The non-transitory computer-readable medium of claim 19 , wherein the instructions, when executed by the processor, cause the processor to:
check an opt-in flag of the third-party node;
when the opt-in flag indicates that the source has opted in to aggregation, receive the globally tuned version of the first-party trained model from the aggregator node; and
when the opt-in flag indicates that the source has not opted in to the aggregation, skip the receiving of the globally tuned version of the first-party trained model from the aggregator node.