US2021110428A1PendingUtilityA1
Click-Through Prediction for Targeted Content
Est. expiryJun 9, 2035(~8.9 yrs left)· nominal 20-yr term from priority
G06N 20/00G06Q 30/0243
66
PatentIndex Score
0
Cited by
0
References
0
Claims
Abstract
In some examples, a computing device includes at least one processor and at least one module, operable by the at least one processor to receive, from a client device of a user, a request for one or more advertisements to display at the client device with a set of messages. The set of messages is associated with the user in a social network messaging service. The at least one module may be further operable to determine a probability that the user will select a candidate advertisement using a machine learning model based on point-wise learning and pair-wise learning. The at least one module may be further operable to determine, based on the probability that the user will select the candidate advertisement, a candidate score for the candidate advertisement, determine that the candidate score satisfies a threshold, and send, for display at the client device, the candidate advertisement.
Claims
exact text as granted — not AI-modified1 . (canceled)
2 . A method for selecting a first targeted content from a set of targeted content to insert into a dynamic timeline of social media content displayed on a user device of a first user during a first session of the first user viewing the timeline, the method comprising:
training, using training data, a classifier to assign a click through probability to each targeted content of the set of targeted content, the training including optimizing a loss function associated with the classifier, wherein the training data includes historical click through data associated with a set of users interacting with historical targeted content during historical sessions associated with the set of users; calculating, with the classifier, a score for each targeted content of the set of targeted content; identifying, the score of the first targeted content as a first score that is greater than the score of every other targeted content of the set of targeted content; and based on the first score meeting a predetermined criterion, sending, by the computing device, the first targeted content to the client device for display in the dynamic timeline during the first session.
3 . The method of claim 2 , wherein the loss function associated with the classifier is a cumulative measure of a loss function associated with each instance of historical click through data.
4 . The method of claim 3 , wherein the loss function associated with each instance of historical click through data is a function of a feature vector associated with that instance of historical click through data.
5 . The method of claim 4 , wherein the loss function associated with each instance of historical click through data is further a function of a label associated with that instance of historical click through data, wherein the label has a first value if the historical targeted content associated with that instance of historical click through data was clicked on by a user associated with that instance of historical click through data, and wherein the label has a second value if the historical targeted content associated with that instance of historical click through data was not clicked on by a user associated with that instance of historical click through data.
6 . The method of claim 5 , further comprising generating the training data dynamically by adding instances of the historical click through data, in real-time and upon generation of each instance of the historical click through data, to the training data, the generating further including setting the label of that instance of the historical click through data to the second value.
7 . The method of claim 6 , further comprising:
receiving an indication of that instance of the historical click through data, which was previously added to the training data with a label set to the second value, being clicked on; and updating the label of that instance of the historical click through data to the first value.
8 . The method of claim 4 , wherein the feature vector includes an advertisement (ad) feature, a user feature, an ad-user interaction feature, and a context feature.
9 . The method of claim 8 , wherein the ad feature include a specification of one or more topics of interest to an advertiser associated with that instance of historical click through data.
10 . The method of claim 8 , wherein the user feature includes a specification of one or more topics of interest to a user of the set of users associated with that instance of historical click through data.
11 . The method of claim 8 , wherein ad-user interaction feature includes a similarity measure, and wherein the similarity measure is based on a measure of similarity between a user profile of a user of the set of users associated with that instance of historical click through data and an advertiser provide of an advertiser associated with that instance of historical click through data.
12 . The method of claim 8 , wherein the context feature includes a specification of a position of the historical targeted content associated with that instance of historical click through data, within a dynamic timeline of a user of the set of users and during the historical session of that user associated with that instance of historical click through data.
13 . The method of claim 8 , wherein the context feature includes a measure of similarity between historical targeted content and other content in the historical session and within a dynamic timeline of a user of the set of users associated with that instance of historical click through data.
14 . The method of claim 13 , wherein the measure of similarity is based on:
a comparison of a bag of words representation of the historical targeted content and of the other content; or a comparison of a word vector representation of the historical targeted content and of the other content.
15 . The method of claim 3 , further comprising instantiating the loss function associated with each instance of historical click through data based on a stochastic gradient descent (SGD) analysis of a feature vector associated with that instance of historical click through data.
16 . The method of claim 2 , wherein the loss function associated with the classifier is:
L
(
w
,
D
)
=
∑
(
y
,
x
)
∈
D
l
(
y
,
f
(
w
,
x
)
)
,
wherein x is a feature vector associated with each instance of historical click through data, y is a binary label associated with the presence or absence of a click for that instance of historical click through data, 1 is a loss function for that instance of historical data, D is a set of all instances of the historical click through data, f is a hypothesis function, and w is one or more parameters of the hypothesis function.
17 . The method of claim 2 , further comprising receiving a request for targeted content from the user device, the request including an indication of the first user refreshing the dynamic timeline of content to initiate the first session.
18 . The method of claim 2 , further comprising identifying the set of targeted content based on prior information associated with the first user.
19 . A non-transitory computer-readable storage medium encoded with instructions for selecting a first targeted content from a set of targeted content to insert into a dynamic timeline of social media content displayed on a user device of a first user during a first session of the first user viewing the timeline, wherein the instructions, when executed, cause one or more processors to:
train, using training data, a classifier to assign a click through probability to each targeted content of the set of targeted content, the training including optimizing a loss function associated with the classifier, wherein the training data includes historical click through data associated with a set of users interacting with historical targeted content during historical sessions associated with the set of users; calculate, with the classifier, a score for each targeted content of the set of targeted content; identify, the score of the first targeted content as a first score that is greater than the score of every other targeted content of the set of targeted content; and based on the first score meeting a predetermined criterion, send, by the computing device, the first targeted content to the client device for display in the dynamic timeline during the first session.
20 . The non-transitory computer-readable storage medium of claim 19 , wherein the loss function associated with the classifier is a cumulative measure of a loss function associated with each instance of historical click through data.
21 . The non-transitory computer-readable storage medium of claim 19 , wherein the loss function associated with each instance of historical click through data is a function of a feature vector associated with that instance of historical click through data.
22 . The non-transitory computer-readable storage medium of claim 21 , wherein the loss function associated with each instance of historical click through data is further a function of a label associated with that instance of historical click through data, wherein the label has a first value if the historical targeted content associated with that instance of historical click through data was clicked on by a user associated with that instance of historical click through data, and wherein the label has a second value if the historical targeted content associated with that instance of historical click through data was not clicked on by a user associated with that instance of historical click through data.
23 . The non-transitory computer-readable storage medium of claim 22 , wherein the instructions further cause the one or more processors to generate the training data dynamically by adding instances of the historical click through data, in real-time and upon generation of each instance of the historical click through data, to the training data, the generating further including setting the label of that instance of the historical click through data to the second value.
24 . The non-transitory computer-readable storage medium of claim 23 , wherein the instructions further cause the one or more processors to:
receive an indication of that instance of the historical click through data, which was previously added to the training data with a label set to the second value, being clicked on; and update the label of that instance of the historical click through data to the first value.
25 . The non-transitory computer-readable storage medium of claim 21 , wherein the feature vector includes an advertisement (ad) feature, a user feature, an ad-user interaction feature, and a context feature.
26 . The non-transitory computer-readable storage medium of claim 19 , wherein the instructions further cause the one or more processors to instantiate the loss function associated with each instance of historical click through data based on a stochastic gradient descent (SGD) analysis of a feature vector associated with that instance of historical click through data.
27 . The non-transitory computer-readable storage medium of claim 19 , wherein the loss function associated with the classifier is:
L
(
w
,
D
)
=
∑
(
y
,
x
)
∈
D
l
(
y
,
f
(
w
,
x
)
)
,
wherein x is a feature vector associated with each instance of historical click through data, y is a binary label associated with the presence or absence of a click for that instance of historical click through data, 1 is a loss function for that instance of historical data, D is a set of all instances of the historical click through data, f is a hypothesis function, and w is one or more parameters of the hypothesis function.
28 . The non-transitory computer-readable storage medium of claim 19 , wherein the instructions further cause the one or more processors to receive a request for targeted content from the user device, the request including an indication of the first user refreshing the dynamic timeline of content to initiate the first session.
29 . A computing device for selecting a first targeted content from a set of targeted content to insert into a dynamic timeline of social media content displayed on a user device of a first user during a first session of the first user viewing the timeline, the computing device comprising:
at least one processor; and at least one non-transitory computer-readable storage medium storing instructions that are executable by the at least one processor to:
train, using training data, a classifier to assign a click through probability to each targeted content of the set of targeted content, the training including optimizing a loss function associated with the classifier, wherein the training data includes historical click through data associated with a set of users interacting with historical targeted content during historical sessions associated with the set of users;
calculate, with the classifier, a score for each targeted content of the set of targeted content;
identify, the score of the first targeted content as a first score that is greater than the score of every other targeted content of the set of targeted content; and
based on the first score meeting a predetermined criterion, send, by the computing device, the first targeted content to the client device for display in the dynamic timeline during the first session.
30 . The computing device of claim 29 , wherein the loss function associated with the classifier is a cumulative measure of a loss function associated with each instance of historical click through data.
31 . The computing device of claim 29 , wherein the loss function associated with each instance of historical click through data is a function of a feature vector associated with that instance of historical click through data.
32 . The computing device of claim 29 , wherein the loss function associated with each instance of historical click through data is further a function of a label associated with that instance of historical click through data, wherein the label has a first value if the historical targeted content associated with that instance of historical click through data was clicked on by a user associated with that instance of historical click through data, and wherein the label has a second value if the historical targeted content associated with that instance of historical click through data was not clicked on by a user associated with that instance of historical click through data.
33 . The computing device of claim 32 , wherein the instructions further cause the at least one processor to generate the training data dynamically by adding instances of the historical click through data, in real-time and upon generation of each instance of the historical click through data, to the training data, the generating further including setting the label of that instance of the historical click through data to the second value.
34 . The computing device of claim 33 , wherein the instructions further cause the at least one processor to:
receive an indication of that instance of the historical click through data, which was previously added to the training data with a label set to the second value, being clicked on; and update the label of that instance of the historical click through data to the first value.
35 . The computing device of claim 31 , wherein the feature vector includes an advertisement (ad) feature, a user feature, an ad-user interaction feature, and a context feature.
36 . The computing device of claim 29 , wherein the loss function associated with the classifier is:
L
(
w
,
D
)
=
∑
(
y
,
x
)
∈
D
l
(
y
,
f
(
w
,
x
)
)
,
wherein x is a feature vector associated with each instance of historical click through data, y is a binary label associated with the presence or absence of a click for that instance of historical click through data, 1 is a loss function for that instance of historical data, D is a set of all instances of the historical click through data, f is a hypothesis function, and w is one or more parameters of the hypothesis function.Join the waitlist — get patent alerts
Track US2021110428A1 — get alerts on status changes and closely related new filings.
We store only your email — no account needed. See our privacy policy.