Method and system for adaptive online updating of ad related models

Patent metadata

Field	Value
Publication number	US-11544740-B2
Application number	US-201715433030-A
Country	US
Kind code	B2
Filing date	Feb 15, 2017
Priority date	Feb 15, 2017
Publication date	Jan 3, 2023
Grant date	Jan 3, 2023

Abstract

The present teaching relates to generating an updated model related to advertisement selection. In one example, a request is obtained for updating a model to be utilized for selecting an advertisement. A plurality of copies of the model is generated. The model is pre-selected based on a performance metric related to advertisement selection. Based on each of the plurality of copies, a candidate model is created by modifying one or more parameters of the copy of the model to create a plurality of candidate models. One of the plurality of candidate models is selected based on the performance metric. The steps of generating, creating, and selecting are repeated until a predetermined condition is met. The model is updated with the latest selected candidate model when the predetermined condition is met.

First claim

We claim:

1. A method, implemented on a machine having at least one processor, storage, and a communication platform connected to a network for generating an updated model related to advertisement selection, the method comprising:
obtaining a request for updating a model to be utilized for selecting an advertisement;
generating a plurality of copies of the model, wherein the model is pre-selected based on a performance metric related to advertisement selection;
creating, for each of the plurality of copies, a candidate model by modifying a value of one or more hyper parameters of the copy of the model based on one or more scale factors to create a plurality of candidate models, wherein the one or more scale factors are determined based on one or more temporal changes indicating a shift from a first time period associated with a first amount of new advertisements to a second time period associated with a second amount, different from the first amount, of new advertisements, and the one or more temporal changes indicate a change of a weight associated with a type of learning of new advertisements associated with learning of the model and are detected from streaming ad-related data;
selecting one of the plurality of candidate models based on the performance metric;
repeating the steps of generating, creating, and selecting until a predetermined condition is met; and
updating the model with the latest selected candidate model when the predetermined condition is met.

2. The method of claim 1 , further comprising:
detecting a divergence of the model when performing the steps of generating, creating, and selecting;
tracking back to a latest cycle of generating, creating, and selecting when the model did not diverge; and
continuing the steps of generating, creating, and selecting from the latest cycle.

3. The method of claim 1 , further comprising:
detecting a convergence of the model when performing the steps of generating, creating, and selecting;
generating an anchor set of hyper parameters that are known to cause the convergence of the model; and
storing the anchor set for future model update.

4. The method of claim 1 , wherein modifying the value of the one or more hyper parameters comprises:
monitoring the one or more temporal changes related to the model; and
generating one or more perturbed values of the one or more hyper parameters based on the one or more scale factors.

5. The method of claim 1 , further comprising:
obtaining ad related data online in real time;
parsing the ad related data according to a timer;
sub-sampling the ad related data;
extracting relevant information from the ad related data based on the model;
determining, in the relevant information, features related to advertisements and users for whom an advertisement is to be selected, to generate processed training data; and
calculating, based on the processed training data, a value for the performance metric for each of the plurality of candidate models, wherein selecting the one of the plurality of candidate models is based on the value for the performance metric for each of the plurality of candidate models.

6. The method of claim 1 , wherein the model is updated online and has been pre-trained offline to determine an initial value of the one or more hyper parameters.

7. The method of claim 1 , wherein the updated model is to be utilized for at least one of:
predicting a probability of click-through rate for an advertisement;
predicting a probability of ad conversion rate for an advertisement;
predicting a total ad revenue for an advertisement; or
predicting user experience for an advertisement.

8. A system having at least one processor, storage comprising computer program instructions stored therein, and a communication platform connected to a network for generating an updated model related to advertisement selection, comprising:
an online training data processor, implemented via the at least one processor executing the computer program instructions, and configured for obtaining a request for updating a model to be utilized for selecting an advertisement;
a hyper parameter set perturbor, implemented via the at least one processor executing the computer program instructions, and configured for:
generating a plurality of copies of the model, wherein the model is pre-selected based on a performance metric related to advertisement selection, and
creating, for each of the plurality of copies, a candidate model by modifying a value of one or more hyper parameters of the copy of the model based on one or more scale factors to create a plurality of candidate models, wherein the one or more scale factors are determined based on one or more temporal changes indicating a shift from a first time period associated with a first amount of new advertisements to a second time period associated with a second amount, different from the first amount, of new advertisements, and the one or more temporal changes indicate a change of a weight associated with a type of learning of new advertisements associated with learning of the model and are detected from streaming ad-related data; and

a hyper parameter tuner, implemented via the at least one processor executing the computer program instructions, and configured for:
selecting one of the plurality of candidate models based on the performance metric, wherein the steps of generating, creating, and selecting are repeated until a predetermined condition is met, and
updating the model with the latest selected candidate model when the predetermined condition is met.

9. The system of claim 8 , wherein:
the hyper parameter tuner comprises a model divergence detector, implemented via the at least one processor executing the computer program instructions, and configured for detecting a divergence of the model when performing the steps of generating, creating, and selecting; and
the system further comprises a back track controller, implemented via the at least one processor executing the computer program instructions, and configured for tracking back to a latest cycle of generating, creating, and selecting when the model did not diverge to continue the steps of generating, creating, and selecting from the latest cycle.

10. The system of claim 8 , wherein the hyper parameter tuner comprises:
a model converge detector, implemented via the at least one processor executing the computer program instructions, and configured for detecting a convergence of the model when performing the steps of generating, creating, and selecting; and
an anchor set updater, implemented via the at least one processor executing the computer program instructions, and configured for:
generating an anchor set of hyper parameters that are known to cause the convergence of the model, and
storing the anchor set for future model update.

11. The system of claim 8 , wherein modifying the value of the one or more hyper parameters comprises:
monitoring the one or more temporal changes related to the model; and
generating one or more perturbed values of the hyper parameters based on the one or more scale factors.

12. The system of claim 8 , wherein the online training data processor comprises:
an ad data parser, implemented via the at least one processor executing the computer program instructions, and configured for obtaining ad related data online in real time and parsing the ad related data according to a timer;
an ad data sub-sampler, implemented via the at least one processor executing the computer program instructions, and configured for sub-sampling the ad related data;
a relevant information extractor, implemented via the at least one processor executing the

Assignees

• Yahoo Holdings Inc
• Yahoo Ad Tech Llc

Inventors

• Kagian Amit
• Aharon Michal
• Somekh Oren Shlomo

Classifications

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06Q30/0254 Primary

  based on statistics · CPC title

• G06N20/20

  Ensemble learning · CPC title

• G06N20/00

  Machine learning · CPC title