US 12,456,045 B2
Embedded learning for response prediction in content item relevance
Seyedmohsen Jamali, Sunnyvale, CA (US); Samaneh Abbasi Moghaddam, Sunnyvale, CA (US); Revant Kumar, Sunnyvale, CA (US); and Vinay Praneeth Boda, Mountain View, CA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Mar. 30, 2019, as Appl. No. 16/370,886.
Prior Publication US 2020/0311543 A1, Oct. 1, 2020
Int. Cl. G06N 3/08 (2023.01); G06N 3/045 (2023.01); G06Q 10/0631 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/045 (2023.01); G06Q 10/063112 (2013.01)] 8 Claims
OG exemplary drawing
 
1. A method comprising:
inserting a first image into a neural network;
the first image is associated with a first historical user interaction with a first content item;
generating, by the neural network, a first image embedding from the first image;
the neural network is trained to learn image embeddings from images;
identifying a second image;
the second image is associated with a possible future user interaction with a second content item;
inserting the second image into the neural network;
generating, by the neural network, a second image embedding from the second image;
based on the first image embedding and the second image embedding, generating a prediction of whether a particular entity will interact with the second content item;
using one or more machine learning techniques to learn weights for a plurality of contextual features while training one or more layers of the neural network;
based on a content request, identifying a plurality of feature values for the plurality of contextual features;
generating the prediction based on the weights and the plurality of feature values; and
by the neural network, machine learning the weights for the plurality of contextual features separately from at least one of the first image embedding or the second image embedding.