Granted Patent B2

US 12,423,576 · App. 17/444,687 · Granted Sep 23, 2025

Inventors: Wenhui Zhang (Beijing, CN); Dianhai Yu (Beijing, CN); Zhihua Wu (Beijing, CN)

Assignee: BEIJING BAIDU NETCOM SCIENCE AND TECHNOLOGY CO., LTD.

G06N3/08G06F8/65G06N3/045G06N3/063G06N5/02G06N20/00G06N20/20

Abstract

Claims (54)

1. A method for updating parameters of a multi-task model, wherein the multi-task model is an advertisement recommendation model, the advertisement recommendation model comprises two subtasks which are predicting a click-through rate and a conversion rate of the advertisement, the method comprising:

obtaining a training sample set, wherein the training sample set comprises a plurality of samples and a task to which each sample belongs;

putting each sample into a corresponding sample queue sequentially according to a task to which each sample belongs;

training a shared network layer in the multi-task model and a target sub-network layer of a task associated with a sample queue with only samples corresponding to the task in the sample queue;

determining a number of the samples in the sample queue reaches a training data requirement;

wherein in response to the determination the number of the samples in the sample queue reaches the training data requirement, determining a model parameter update gradient corresponding to the task associated with the sample queue; and

updating parameters of the shared network layer and the target sub-network layer in a parameter server according to the model parameter update gradient;

wherein the training is a supervised training scenario, wherein a sample comprises training data and annotation data corresponding to the training data, the training data is feature data of an advertisement comprising the type, the duration, or the label, and the annotation data comprises user operation data comprising a click-through rate, viewing time, a number of favorites, a number of reposts, a number of shares, a conversion rate;

wherein putting each sample into the corresponding sample queue sequentially according to a task to which each sample belongs comprises:

determining a task label corresponding to each sample according to a task to which each sample belongs; and

putting each sample sequentially into a sample queue corresponding to the task label according to the task label corresponding to each sample;

wherein updating the parameters of the shared network layer and the target sub-network layer in the parameter server according to the model parameter update gradient comprises:

determining weights of the task associated with the sample queue;

updating the parameters of the target sub-network layer according to the model parameter update gradient;

determining an update gradient of the shared network layer according to the weights and the model parameter update gradient; and

updating the parameters of the shared network layer according to the update gradient of the shared network layer.

2. The method of claim 1 , wherein putting each sample sequentially into the sample queue corresponding to the task label according to the task label corresponding to each sample comprises:

determining, in the case that a sample corresponds to the plurality of task labels, that the sample queue corresponding to each task label of a plurality of task labels contains the sample.

3. An apparatus for updating parameters of a multi-task model, wherein the multi-task model is an advertisement recommendation model, the advertisement recommendation model comprises two subtasks which are predicting a click-through rate and a conversion rate of the advertisement, the method comprising:

one or more processors;

a memory storing instructions executable by the one or more processors;

wherein the one or more processors are configured to:

obtain a training sample set, wherein the training sample set comprises a plurality of samples and a task to which each sample belongs;

put each sample into a corresponding sample queue sequentially according to a task to which each sample belongs;

train a shared network layer in the multi-task model and a target sub-network layer of a task associated with a sample queue with only samples corresponding to the task in the sample queue;

determine a number of the samples in the sample queue reaches a training data requirement;

wherein in response to the determination the number of the samples in the sample queue reaches the training data requirement, determine a model parameter update gradient corresponding to the task associated with the sample queue; and

update parameters of the shared network layer and the target sub-network layer in a parameter server according to the model parameter update gradient;

wherein the training is a supervised training scenario, wherein a sample comprises training data and annotation data corresponding to the training data, the training data is feature data of an advertisement comprising the type, the duration, or the label, and the annotation data comprises user operation data comprising a click-through rate, viewing time, a number of favorites, a number of reposts, a number of shares, a conversion rate;

wherein the one or more processors are configured to:

determine a task label corresponding to each sample according to a task to which each sample belongs; and

put each sample sequentially into a sample queue corresponding to the task label according to the task label corresponding to each sample;

determine weights of the task associated with the sample queue;

update the parameters of the target sub-network layer according to the model parameter update gradient;

determine an update gradient of the shared network layer according to the weights and the model parameter update gradient; and

update the parameters of the shared network layer according to the update gradient of the shared network layer.

4. The apparatus of claim 3 , wherein the one or more processors are configured to:

determine, in the case that a sample corresponds to the plurality of task labels, that the sample queue corresponding to each task label of a plurality of task labels contains the sample.

5. A non-transitory computer-readable storage medium storing computer instructions, wherein when the computer instructions are executed by a computer, the computer is caused to perform a method for updating parameters of a multi-task model, wherein the multi-task model is an advertisement recommendation model, the advertisement recommendation model comprises two subtasks which are predicting a click-through rate and a conversion rate of the advertisement, and the method comprises:

obtaining a training sample set, wherein the training sample set comprises a plurality of samples and a task to which each sample belongs;

putting each sample into a corresponding sample queue sequentially according to a task to which each sample belongs;

training a shared network layer in the multi-task model and a target sub-network layer of a task associated with a sample queue with only samples corresponding to the task in the sample queue;

determining a number of the samples in the sample queue reaches a training data requirement;

wherein in response to the determination the number of the samples in the sample queue reaches the training data requirement, determining a model parameter update gradient corresponding to the task associated with the sample queue; and

update parameters of the shared network layer and the target sub-network layer in a parameter server according to the model parameter update gradient;

wherein the training is a supervised training scenario, wherein a sample comprises training data and annotation data corresponding to the training data, the training data is feature data of an advertisement comprising the type, the duration, or the label, and the annotation data comprises user operation data comprising a click-through rate, viewing time, a number of favorites, a number of reposts, a number of shares, a conversion rate;

wherein putting each sample into the corresponding sample queue sequentially according to a task to which each sample belongs comprises:

determining a task label corresponding to each sample according to a task to which each sample belongs; and

putting each sample sequentially into a sample queue corresponding to the task label according to the task label corresponding to each sample;

wherein updating the parameters of the shared network layer and the target sub-network layer in the parameter server according to the model parameter update gradient comprises:

determining weights of the task associated with the sample queue;

updating the parameters of the target sub-network layer according to the model parameter update gradient;

determining an update gradient of the shared network layer according to the weights and the model parameter update gradient; and

updating the parameters of the shared network layer according to the update gradient of the shared network layer.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (19)