Privacy-safe frequency distribution of geo-features for mobile devices
A mobile device's location and identifier corresponding to the mobile device is received. The mobile device's location is mapped to a plurality of landmarks proximate to the mobile device's location. The proximate landmarks are stored in association with the mobile device's identifier in a geo data store, and the received location data is then discarded. These steps are iterated over time to build up a data store that can be represented as a frequency distribution of the landmarks that surround mobile devices. Such a frequency distribution can be built for each of a plurality of mobile devices, without maintaining records of any mobile device's location.
1. A method comprising:
generating, by a server, a meta model, comprising:
for each of a plurality of mobile devices, for each of a plurality of locations:
receiving a location and an identifier corresponding to a respective mobile device;
mapping the location to a plurality of proximate landmarks;
storing the proximate landmarks in association with the identifier; and
discarding the received location;
featurizing the plurality of proximate landmarks by frequency to generate a set of frequency features;
training a geo model for a campaign using the frequency features; and
combining the geo model with a pixel model to generate a meta model; and
automatically generating, by the server, a response to a bid request, comprising:
receiving, from an ad exchange, the bid request and a target identifier corresponding to an advertising opportunity;
receiving a consumption history and record of proximate landmarks associated with the target identifier; and
applying the meta model to the consumption history and record of proximate landmarks associated with the target identifier to determine a value of the advertising opportunity to an advertising campaign;
returning, to the ad exchange, a bid amount based on the determined value;
wherein the time passage between receiving the bid request and returning the bid amount is a fraction of a second.
2. The method of claim 1 , wherein the proximate landmarks comprise categories of businesses.
3. The method of claim 1 , wherein featurizing the plurality of proximate landmarks by frequency further comprises, for each landmark category, dividing a count in the respective landmark category by a total number of observations of a location of the mobile device.
4. The method of claim 1 , wherein training the geo model further comprises training the geo model to identify geo features that are most strongly correlated to a status of converter or non-converter, respectively.
5. The method of claim 1 , wherein combining the geo model with the pixel model comprises training the meta model to determine the respective weight to apply to each feature in the meta model.
6. The method of claim 1 , wherein returning the bid amount further comprises sending an ad redirect to an ad exchange.
7. A non-transitory computer-readable storage medium executing computer program instructions, the computer program instructions comprising instructions for:
generating, by a server, a meta model, comprising:
for each of a plurality of mobile devices, for each of a plurality of locations:
receiving a location and an identifier corresponding to a respective mobile device;
mapping the location to a plurality of proximate landmarks;
storing the proximate landmarks in association with the identifier; and
discarding the received location;
featurizing the plurality of proximate landmarks by frequency to generate a set of frequency features;
training a geo model for a campaign using the frequency features; and
combining the geo model with a pixel model to generate a meta model; and
automatically generating, by the server, a response to a bid request, comprising:
receiving, from an ad exchange, the bid request and a target identifier corresponding to an advertising opportunity;
receiving a consumption history and record of proximate landmarks associated with the target identifier; and
applying the meta model to the consumption history and record of proximate landmarks associated with the target identifier to determine a value of the advertising opportunity to an advertising campaign;
returning, to the ad exchange, a bid amount based on the determined value;
wherein the time passage between receiving the bid request and returning the bid amount is a fraction of a second.
8. The medium of claim 7 , wherein the proximate landmarks comprise categories of businesses.
9. The medium of claim 7 , wherein the instructions for featurizing the plurality of proximate landmarks by frequency further comprise instructions for dividing, for each landmark category, a count in the respective landmark category by a total number of observations of a location of the mobile device.
10. The medium of claim 7 , wherein the instructions for training the geo model further comprise instructions for training the geo model to identify geo features that are most strongly correlated to a status of converter or non-converter, respectively.
11. The medium of claim 7 , wherein the instructions for combining the geo model with the pixel model further comprise instructions for training the meta model to determine the respective weight to apply to each feature in the meta model.
12. The medium of claim 7 , wherein returning the bid amount further comprises sending an ad redirect to an ad exchange.
13. A system comprising:
a processor;
a computer-readable storage medium storing processor-executable computer program instructions, the instructions comprising instructions for:
generating, by a server, a meta model, comprising:
for each of a plurality of mobile devices, for each of a plurality of locations:
receiving a location and an identifier corresponding to a respective mobile device;
mapping the location to a plurality of proximate landmarks;
storing the proximate landmarks in association with the identifier; and
discarding the received location;
featurizing the plurality of proximate landmarks by frequency to generate a set of frequency features;
training a geo model for a campaign using the frequency features; and
combining the geo model with a pixel model to generate a meta model; and
automatically generating, by the server, a response to a bid request, comprising:
receiving, from an ad exchange, the bid request and a target identifier corresponding to an advertising opportunity;
receiving a consumption history and record of proximate landmarks associated with the target identifier; and
applying the meta model to the consumption history and record of proximate landmarks associated with the target identifier to determine a value of the advertising opportunity to an advertising campaign;
returning, to the ad exchange, a bid amount based on the determined value;
wherein the time passage between receiving the bid request and returning the bid amount is a fraction of a second.
14. The system of claim 13 , wherein the proximate landmarks comprise categories of businesses.
15. The system of claim 13 , wherein the instructions for featurizing the plurality of proximate landmarks by frequency further comprise instructions for dividing, for each landmark category, a count in the respective landmark category by a total number of observations of a location of the mobile device.
16. The system of claim 13 , wherein the instructions for training the geo model further comprise instructions for training the geo model to identify geo features that are most strongly correlated to a status of converter or non-converter, respectively.
17. The system of claim 13 , wherein the instructions for combining the geo model with the pixel model further comprise instructions for training the meta model to determine the respective weight to apply to each feature in the meta model.
18. The system of claim 13 , wherein returning the bid amount further comprises sending an ad redirect to an ad exchange.