US 12,456,036 B2
Deep embedding learning models with mimicry effect
Yuan Sun, Redmond, WA (US); Ye Tu, San Carlos, CA (US); Ying Han, Sunnyvale, CA (US); Chun Lo, Mountain View, CA (US); Shaunak Chatterjee, Sunnyvale, CA (US); and Vrishti Gulati, Milpitas, CA (US)
Assigned to MICROSOFT TECHNOLOGY LICENSING, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Dec. 20, 2021, as Appl. No. 17/556,218.
Prior Publication US 2023/0196070 A1, Jun. 22, 2023
Int. Cl. G06N 20/20 (2019.01); G06N 3/045 (2023.01); G06N 3/047 (2023.01)
CPC G06N 3/045 (2023.01) [G06N 3/047 (2023.01); G06N 20/20 (2019.01)] 20 Claims
OG exemplary drawing
 
10. A method comprising:
accessing training data corresponding to a plurality of users, the training data including information about user interaction with items of a plurality of item types in an online network;
splitting the training data into a first treatment group, a second treatment group, and a common control group, the first treatment group including training data corresponding to users who interacted with items of a first item type, the second treatment group including training data corresponding to users who interacted with items of a second item type, and the common control group including training data corresponding to users who did not interact with items of either the first item type or the second item type;
feeding the first treatment group and the common control group into a machine learning algorithm to train a first mimicry model to output a first mimicry score indicative of a mimicry effect on a user presented with an item of the first item type, the machine learning algorithm being a counterfactual regression neural network, wherein a mimicry score comprises a delta between a probability of a user interacting with an item of a corresponding item type if the user was previously exposed to items of the corresponding item type and a probability of the user interacting with an item of the corresponding item type if the user was not previously exposed to items of the corresponding item type, the delta comprising a positive or negative correlation to interactions with previously exposed items of the corresponding item type;
feeding the second treatment group and the common control group into the machine learning algorithm to train a second mimicry model to output a second mimicry score indicative of a mimicry effect on a user presented with an item of the second item type;
ranking, by a ranking model, feed objects for a user based on the first mimicry score and the second mimicry score, the feed objects comprising items of the first item type and items of the second item type; and
updating a feed of the user with a feed object based on the ranking of feed objects.