WO2023116138A1 - 多任务模型的建模方法、推广内容处理方法及相关装置 - Google Patents

多任务模型的建模方法、推广内容处理方法及相关装置 Download PDF

Info

Publication number
WO2023116138A1
WO2023116138A1 PCT/CN2022/124765 CN2022124765W WO2023116138A1 WO 2023116138 A1 WO2023116138 A1 WO 2023116138A1 CN 2022124765 W CN2022124765 W CN 2022124765W WO 2023116138 A1 WO2023116138 A1 WO 2023116138A1
Authority
WO
WIPO (PCT)
Prior art keywords
task
tasks
model
promotion
content
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2022/124765
Other languages
English (en)
French (fr)
Inventor
吴寅初
佘琪
王长虎
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Youzhuju Network Technology Co Ltd
Original Assignee
Beijing Youzhuju Network Technology Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Youzhuju Network Technology Co Ltd filed Critical Beijing Youzhuju Network Technology Co Ltd
Publication of WO2023116138A1 publication Critical patent/WO2023116138A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0251—Targeted advertisements
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q30/00—Commerce
    • G06Q30/02—Marketing; Price estimation or determination; Fundraising
    • G06Q30/0241—Advertisements
    • G06Q30/0277—Online advertisement

Definitions

  • the present disclosure belongs to the technical field of artificial intelligence, and specifically relates to a modeling method of a multi-task model, a promotion content processing method, a device, a device, a computer-readable storage medium, and a computer program product.
  • the application scenarios of artificial intelligence technology are becoming more and more extensive.
  • promotional content such as an advertisement
  • the conversion rate of the promotional content can be predicted based on artificial intelligence technology, and then the promotional content can be pushed to users based on the conversion rate.
  • multi-task learning (muti task learning, MTL) is usually used to build a multi-task model.
  • MTL multi-task learning
  • multiple tasks are selected based on artificial subjective experience for multi-task learning, so that the learned features can be shared between different tasks, thereby improving the learning efficiency and generalization ability of the multi-task model.
  • the purpose of the present disclosure is to provide a modeling method of a multi-task model, a promotion content processing method, a device, a device, a computer-readable storage medium, and a computer program product, which can improve the learning efficiency of a multi-task model.
  • the present disclosure provides a modeling method of a multi-task model, including:
  • the initial task set includes: the conversion rate of promotional content, the playing duration of the promotional content, and the presentation type of the promotional content and at least two of the information on promotion objects in the promotion content;
  • a sample of a multi-task model is generated, and the sample of the multi-task model is used for model training to obtain the multi-task model.
  • the present disclosure provides a method for processing promotional content, including:
  • the inference result of the multi-task model is obtained;
  • the multi-task model is generated based on the features corresponding to each task in the related task set. samples, the set of related tasks is obtained based on the mutual information between different tasks, and the different tasks are the tasks in the initial task set constructed by the features used to construct the tasks;
  • the reasoning results include the conversion rate of the promotion content , the playback duration of the promotion content, the presentation type of the promotion content, or the information of the promotion object in the promotion content;
  • the promotion strategy for the promotion content is adjusted.
  • the present disclosure provides a modeling device for a multi-task model, including:
  • An acquisition module configured to acquire features for constructing tasks, and construct an initial task set according to the features for constructing tasks;
  • the initial task set includes: the conversion rate of the promotion content, the playing time of the promotion content, the At least two of the presentation type of the promotion content and the information of the promotion object in the promotion content;
  • a mutual information determination module configured to determine mutual information between different tasks in the initial task set
  • a related task determination module configured to obtain a set of related tasks according to the mutual information between the different tasks, and the mutual information of the tasks included in the set of related tasks satisfies a first preset condition
  • a training module configured to generate samples of a multi-task model according to features corresponding to each task in the set of related tasks, and use the samples of the multi-task model to perform model training to obtain the multi-task model.
  • the present disclosure provides a device for processing promotional content, which is characterized in that it includes:
  • An acquisition module configured to acquire the attributes of the user's behavior on the promotion content
  • a reasoning module configured to obtain a reasoning result of the multi-task model according to the attribute of the user's behavior on the promotional content and the multi-task model; the multi-task model is generated based on the features corresponding to each task in the related task set The sample of the multi-task model is obtained, and the related task set is obtained based on mutual information between different tasks, and the different tasks are tasks in the initial task set constructed by the features used to construct the task; the reasoning results include The conversion rate of the promotion content, the playing duration of the promotion content, the presentation type of the promotion content, or the information of the promotion object in the promotion content;
  • a processing module configured to adjust a promotion strategy for the promotion content according to the reasoning result.
  • the present disclosure provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processing device, the steps of the method described in any one of the first aspect or the second aspect of the present disclosure are implemented.
  • an electronic device including:
  • a processing device configured to execute the computer program in the storage device to implement the steps of the method described in any one of the first aspect or the second aspect of the present disclosure.
  • the present disclosure provides a computer program product including instructions, which, when run on a device, cause the device to execute the method described in any implementation manner of the first aspect or the second aspect above.
  • the present disclosure has the following advantages:
  • the present disclosure provides a modeling method of a multi-task model.
  • the features used for constructing tasks are obtained first, an initial task set is constructed based on the features used for constructing tasks, and then the relationship between different tasks in the initial task set is determined.
  • Mutual information Based on the mutual information, the initial task set is screened to obtain a set of related tasks with strong correlation.
  • samples of the multi-task model are generated to train the multi-task model.
  • tasks in the related task set obtained after screening the initial task set in the present disclosure have a stronger correlation. In this way, during the multi-task learning process, the learning efficiency of the multi-task model can be improved.
  • FIG. 1 is a flow chart of a modeling method for a multi-task model provided by an embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of obtaining a set of related tasks provided by an embodiment of the present disclosure
  • FIG. 3 is a schematic diagram of a multi-task model provided by an embodiment of the present disclosure.
  • FIG. 4 is a flowchart of another modeling method for a multi-task model provided by an embodiment of the present disclosure
  • FIG. 5 is a flow chart of a method for processing promotional content provided by an embodiment of the present disclosure
  • FIG. 6 is a schematic diagram of a modeling device for a multi-task model provided by an embodiment of the present disclosure
  • FIG. 7 is a schematic diagram of an apparatus for processing promotional content provided by an embodiment of the present disclosure.
  • FIG. 8 is a schematic structural diagram of an electronic device provided by an embodiment of the present disclosure.
  • first and second in the embodiments of the present disclosure are used for description purposes only, and cannot be interpreted as indicating or implying relative importance or implicitly indicating the quantity of indicated technical features. Thus, a feature defined as “first” and “second” may explicitly or implicitly include one or more of these features.
  • a multi-task model refers to a model constructed based on multi-task learning.
  • the core of multi-task learning is that multiple tasks are trained in parallel and share the learned features with each other.
  • each content can be abstracted into a task, such as predicting conversion rate, predicting potential users, predicting whether users click on the promotion content, etc.
  • a multi-task model can be constructed based on multi-task learning to predict multiple contents.
  • an embodiment of the present disclosure provides a modeling method for a multi-task model, which can be executed by an electronic device.
  • An electronic device may be a server.
  • the server may be a cloud server, for example, a central server in a central cloud computing cluster, or an edge server in an edge cloud computing cluster.
  • the server may also be a server in a local data center.
  • An on-premises data center refers to a data center directly controlled by the user.
  • the modeling method of the multi-task model includes: the electronic device acquires features for constructing tasks, constructs an initial task set according to the features for constructing tasks, and then determines mutual information between different tasks in the initial task set. Based on the mutual information between different tasks, a set of related tasks is obtained, and the mutual information in the set of related tasks satisfies a first preset condition. According to the characteristics corresponding to each task in the related task set, samples of the multi-task model are generated, and the samples of the multi-task model are used for model training to obtain the multi-task model.
  • the electronic device first filters the initial task set, and the tasks in the obtained related task set have a strong correlation. Compared with multiple tasks selected purely relying on subjective experience, the correlation between multiple tasks corresponding to the multi-task model is improved. In this way, during the multi-task learning process, the learning efficiency of the multi-task model can be improved.
  • the multi-task model obtained by using the multi-task model modeling method provided by the embodiments of the present disclosure can be applied to various scenarios.
  • the multi-task model can be used to predict the conversion rate of the promotional content and the playing duration of the promotional content (the time elapsed from the start of the promotional content until the user closes the promotional content).
  • the electronic device may input the attribute of the user's behavior on the promotion content into the multi-task model, and then obtain the above reasoning result. Then the electronic device can adjust the promotion strategy of the promotion content based on the above reasoning result.
  • the conversion rate of the promotional content is greater than or equal to the conversion rate threshold and the playing time of the promotional content is greater than or equal to the duration threshold, the number of promotions of the promotional content is increased.
  • the conversion rate of the promotional content is less than the conversion rate threshold and the playing time of the promotional content is less than the duration threshold, the number of promotions of the promotional content is reduced.
  • this figure is a flowchart of a modeling method for a multi-task model provided by an embodiment of the present disclosure, the method includes:
  • S101 The electronic device acquires features for constructing tasks, and constructs an initial task set according to the features for constructing tasks.
  • An electronic device may obtain initial data, which includes a plurality of characteristics.
  • the initial data may include features such as the conversion rate of the promotional content, the playing time of the promotional content, the presentation type of the promotional content, the information of the promotional objects in the promotional content, and the generation time of the promotional content.
  • Electronic devices can acquire features for building tasks based on initial data.
  • the electronic device may receive a plurality of features configured by a developer as features for building a task.
  • the electronic device can present multiple candidate features to the user through a display device (such as a display screen), such as the features included in the above-mentioned initial data, and then according to the developer's selection operation on the multiple candidate features, the feature selected by the developer As a feature for building tasks.
  • the electronic device constructs an initial task set based on the features for constructing tasks.
  • S102 The electronic device determines mutual information between different tasks in the initial task set.
  • MI Mutual information
  • the electronic device can separately calculate the mutual information between different tasks in the initial task set.
  • the electronic device can calculate the mutual information between different tasks based on the characteristics of the tasks in the initial task set.
  • the electronic device may calculate the mutual information between different tasks in the initial task set based on the following formula:
  • I(X; Y) is the mutual information between task X and task Y in the initial task set; x is the feature corresponding to task X, y is the feature corresponding to task Y; p(x, y) is the task X and task Y The probability that task Y occurs at the same time, p(x) is the probability of task X occurring, and p(y) is the probability of task Y occurring.
  • S103 The electronic device obtains a related task set according to the mutual information between different tasks, and the mutual information of the tasks included in the related task set satisfies a first preset condition.
  • the electronic device may filter the initial task set based on mutual information between different tasks in the initial task set, and then obtain a related task set.
  • the mutual information of tasks included in the set of related tasks satisfies a first preset condition.
  • the first preset condition may be that the mutual information between different tasks is greater than or equal to a mutual information threshold.
  • the electronic device may receive the mutual information threshold configured by the developer, and determine the related task set based on the configured mutual information threshold, so that the developer can adjust the tasks in the related task set within a certain range. .
  • the mutual information threshold may be a value configured by a developer, or may be a default value.
  • the mutual information threshold may be 0.1, 0.2, and so on.
  • S104 The electronic device generates a sample of the multi-task model according to the features corresponding to each task in the related task set, and uses the samples of the multi-task model to perform model training to obtain the multi-task model.
  • the electronic device After determining the related task set, the electronic device can build a multi-task model based on the tasks in the related task set. Since the mutual information between the tasks in the related task set satisfies the first preset condition, it indicates that the tasks in the related task set are highly correlated. Therefore, after constructing a multi-task model based on tasks with strong correlation, the learning efficiency of the multi-task model is higher.
  • the electronic device can generate samples of the multi-task model based on features corresponding to each task in the related task set.
  • the samples of the multi-task model include feature vectors of the samples and labels of the feature vectors of the samples.
  • the initial data includes multiple features, and the features corresponding to the tasks in the related task set are some of the multiple features of the initial data. Based on this, the electronic device can use the feature corresponding to the task in the related task set as the label of the feature vector of the sample, remove the feature corresponding to the task in the above related task set from the initial data, and use the initial data obtained as the feature vector of the sample.
  • the initial data includes feature 1, feature 2, feature 3, feature 4, and feature 5.
  • task 1 corresponds to feature 1
  • task 2 corresponds to feature 2.
  • the electronic device can use feature 1 as the label of the feature vector of the sample used for monitoring task 1, and feature 2 as the label of the feature vector of the sample used for monitoring task 2, remove feature 1 and feature 2 from the initial data, and obtain the following:
  • the initial data of feature 3, feature 4, and feature 5 are used as feature vectors of samples of the multi-task model.
  • the electronic device can use the samples of the multi-task model to perform model training to obtain the multi-task model.
  • FIG. 3 is a schematic diagram of a multi-task model provided by an embodiment of the present disclosure.
  • the multi-task model includes a task exclusive network 310 , an output network 330 , and a shared network 320 corresponding to multiple tasks in the related task set.
  • the task-exclusive network 310 can be a deep neural network (deep neural networks, DNN), a convolutional neural network (convolutional neural network, CNN) or a self-attention network;
  • the shared network 320 can also be a deep neural network, a volume product neural network or self-attention network.
  • the present disclosure does not specifically limit the types of the task exclusive network 310 and the shared neural network 320 .
  • the shared network 320 and the task-exclusive network 310 can be constructed by using the convolutional neural network.
  • a shared network 320 and a task-exclusive network 310 can be constructed using a deep neural network.
  • the self-attention network can be used to construct the shared network 320 and the task-exclusive network 310 .
  • the output network 330 can use activation functions such as sigmoid and relu to obtain the output corresponding to each task (for example, classification value and regression value).
  • the electronic device may input the feature vector of the sample of the multi-task model into the shared network 320 to obtain the shared component, and then input the shared component to the task-exclusive network of each task in the relevant task set to obtain the eigenvector of each task-exclusive network. output. Then the electronic device can determine the loss value based on the label value of the feature vector of the sample of the multi-task model and the output of the task exclusive network, and update the weight of the task exclusive network based on the loss value to perform model training, and then obtain the multi-task model.
  • the embodiments of the present disclosure provide a modeling method for a multi-task model. Compared with multiple tasks selected purely relying on subjective experience, the related task set obtained after the electronic device screens the initial task set The correlation between tasks is strong. In this way, during the multi-task learning process, the learning efficiency of the multi-task model can be improved. Further, it meets the business needs in various scenarios and realizes more accurate prediction of multiple contents.
  • the embodiment of the present disclosure also provides a modeling method of a multi-task model. Based on the embodiment shown in FIG. The number of positive samples for each task group. As shown in Figure 4, the modeling method of the multi-task model also includes:
  • S401 The electronic device aggregates tasks satisfying a second preset condition in a set of related tasks.
  • the second preset condition may be a plurality of tasks in the set of related tasks whose task correlation is greater than a correlation threshold.
  • the electronic device may aggregate the tasks in the related task set based on the second preset condition, and the aggregated related task set includes multiple task groups.
  • the electronic device can obtain the correlation between the outputs of the task exclusive network 310 corresponding to multiple tasks in the related task set, and then based on the correlation between the outputs of the task exclusive network 310 , the related task set Multiple tasks are aggregated. Specifically, the electronic device may calculate the correlation between the outputs of the task exclusive network 310 corresponding to multiple tasks in the related task set based on the following formula:
  • r mn is the correlation between task m and task n in the related task set;
  • O im is the output of sample i in the task exclusive network 310 corresponding to task m, is the mean value of the output of the task exclusive network 310 corresponding to task m to all samples;
  • O in is the output of sample i in the task exclusive network 310 corresponding to task n, is the mean value of the output of all samples of the task exclusive network 310 corresponding to task n, and K is the total number of samples.
  • the electronic device may compare the correlation r mn between task m and task n to a correlation threshold. When r mn ⁇ the correlation threshold, the task m and the task n are aggregated into one group, so that the electronic device can obtain the aggregated related task set including multiple task groups.
  • the electronic device determines multiple target tasks according to the number of positive samples of tasks in the task group.
  • the aggregated set of related tasks includes multiple task groups. Take multiple task groups including task group G1 and task group G2 as an example, where task group G1 includes task T1, task T2, and task T3, the number of positive samples for task T1 is 100, the number of positive samples for task T2 is 200, and the number of positive samples for task T2 is 200.
  • the number of positive samples of T3 is 300; the task group G2 includes task T4, task T5 and task T6, the number of positive samples of task T4 is 200, the number of positive samples of task T5 is 400, and the number of positive samples of task T6 is 500. It can be seen that the sum of the number of positive samples in the task group G1 is 600, and the sum of the number of positive samples in the task group G2 is 1100.
  • the electronic device may use a sum of 600 positive samples of the task group G1 as the positive sample number threshold.
  • the electronic device can sort task T4, task T5, and task T6 in task group G2, such as adding up one by one in order of the number of positive samples from small to large, until the number of positive samples of tasks in task group G2 The sum is greater than or equal to the above positive sample number threshold 600.
  • the electronic device may determine that when the sum of the number of positive samples of tasks in the task group is greater than or equal to the threshold of the number of positive samples, multiple tasks in the task group are multiple target tasks. In this way, after the electronic device sums the number of positive samples of task T4 and the number of positive samples of task T5, the sum result is equal to the threshold of the number of positive samples, and the electronic device can regard task T4 and task T5 as multiple tasks in the task group. target task.
  • the electronic device may randomly select the number of positive samples of multiple tasks in the task group G2 to add up until the number of positive samples of multiple tasks is greater than or equal to the threshold of the number of positive samples, and then the number of randomly selected tasks Determined as multiple target tasks.
  • the electronic device may sum the number of positive samples of task T5 and task T6 in the task group G2, and the sum result is greater than the threshold of the number of positive samples. In this way, the electronic device can determine the task T5 and the task T6 as multiple target tasks.
  • the sum of the number of positive samples of task T4 and task T5 in task group G2 is greater than or equal to the threshold of the number of positive samples means: when subtracting the number of positive samples of either task T4 or task T5 in task group G2 When , the sum of the number of positive samples of the remaining tasks will be less than the threshold of the number of positive samples.
  • the electronic device may remove tasks other than the target task in the task group. Taking the multiple target tasks in the task group G2 as task T4 and task T5 as an example, the electronic device may remove task T6 from the task group G2.
  • the electronic device may generate samples of the multi-task model based on features corresponding to multiple target tasks in the task group.
  • the target tasks included in task group G2 as task T4 and task T5 as an example for this task group G2, the label of the feature vector of the positive sample corresponding to task T4 is used as the label of the task group G2, and the positive sample corresponding to task T5 is The label of the feature vector of the sample is also used as the label of the task group G2.
  • the electronic device can balance the number of positive samples included in multiple task groups to reduce the reduction in the learning efficiency of the multi-task model due to large differences in the number of positive samples.
  • the modeling method of the multi-task model analyzes the statistical information of large-scale data containing multiple characteristics, determines the set of related tasks based on correlation, and realizes the aggregation of highly related tasks.
  • Using highly relevant tasks for multi-task model training can reduce negative transfer and improve the learning efficiency of multi-task models.
  • the structure of the multi-task model is constructed based on the determined set of related tasks, and the task aggregation with model generalization information is completed by combining the Pearson coefficient.
  • the number of positive samples corresponding to the task group with the smallest number of positive samples is used as the positive sample number threshold to screen and combine the tasks of each group, so as to achieve the balance of positive samples in different tasks.
  • the number of positive samples between different tasks can be completely proportional, so that the multi-task model can complete the multi-task learning process without bias.
  • the embodiment of the present disclosure also provides a promotion content processing method, as shown in FIG. 5 , the promotion content processing method includes:
  • S501 The electronic device acquires an attribute of a user's behavior on promotional content.
  • the attributes of the user's behavior on the promotional content may include the duration of the user's viewing of the promotional content, whether the user clicks on the promotional content, the presentation type of the promotional content clicked by the user (such as video type, picture type, etc.), whether the user is Promote content conversion, etc.
  • the electronic device needs to obtain the user's authorization in advance, and the electronic device can only obtain the user's behavior on the promotional content after obtaining the user's authorization to use the corresponding data (such as the above-mentioned attributes of the user's behavior on the promotional content). attributes and other data.
  • S502 The electronic device obtains an inference result of the multi-task model according to the attribute of the user's behavior on the promotion content and the multi-task model.
  • the multi-task model is obtained based on samples of the multi-task model generated by features corresponding to each task in the related task set.
  • the related task set is obtained based on the mutual information between different tasks, and the different tasks are the tasks in the initial task set constructed by the features used to construct the tasks.
  • the reasoning result includes a conversion rate of the promotion content, a playing time of the promotion content, a presentation type of the promotion content, or information of promotion objects in the promotion content.
  • S503 The electronic device adjusts a promotion strategy for the promotion content according to the reasoning result.
  • the electronic device when the inference result shows that the conversion rate of the promotion content is greater than the conversion rate threshold, and the playing time of the promotion content is greater than the duration threshold, the electronic device increases the number of promotions of the promotion content. When the conversion rate of the promotion content is less than the conversion rate threshold and the playing time of the promotion content is less than the time threshold, the electronic device reduces the number of promotions of the promotion content.
  • the presentation type of the promotion content is a preset type (such as a video type)
  • the information of the promotion object in the promotion content is preset information (such as indicating that the promotion object is a game, a virtual item or a physical object)
  • the electronic The device increases the promotion frequency of the promotion content. In this way, the ineffective delivery of promotional content is reduced, and the waste of resources is reduced.
  • Fig. 6 is a schematic diagram of a multi-task model modeling device according to an exemplary disclosed embodiment. As shown in Fig. 6, the multi-task model modeling device 600 includes:
  • the obtaining module 601 is used to obtain the features used to construct the task, and construct an initial task set according to the features used to construct the task; the initial task set includes: the conversion rate of the promotion content, the playing duration of the promotion content, the At least two of the presentation type of the promotion content and the information of the promotion object in the promotion content;
  • a related task determining module 603, configured to obtain a related task set according to mutual information between the different tasks, and the mutual information of the tasks included in the related task set satisfies a first preset condition
  • the training module 604 is configured to generate a sample of a multi-task model according to features corresponding to each task in the set of related tasks, and use the samples of the multi-task model to perform model training to obtain the multi-task model.
  • the related task determining module 603 is further configured to aggregate tasks satisfying a second preset condition in the set of related tasks.
  • the related task determination module 603 is specifically configured to determine the correlation between the outputs of the task exclusive network corresponding to multiple tasks in the related task set; Multiple tasks are aggregated.
  • the relevant task determination module 603 is further configured to determine multiple target tasks for each task group according to the number of positive samples of tasks in the task group, and one of the number of positive samples of the multiple target tasks is The sum is greater than the threshold of the number of positive samples, and the threshold of the number of positive samples is the minimum value of the sum of the number of positive samples in each task group;
  • the training module 604 is specifically configured to generate a sample of a multi-task model according to the features corresponding to the target task in the task group.
  • the relevant task determination module 603 is specifically configured to add up the number of positive samples of the tasks in the task group from small to large, until the sum of the number of positive samples of the tasks in the task group greater than or equal to the threshold of the number of positive samples; when it is determined that the sum of the number of positive samples of tasks in the task group is greater than or equal to the threshold of the number of positive samples, the multiple tasks in the task group are multiple target tasks.
  • the training module 604 is specifically configured to input the feature vectors of the samples of the multi-task model into the shared network to obtain shared components; input the shared components to the task exclusives of each task in the related task set The network obtains the output corresponding to the task-exclusive network of each task; and trains the multi-task model according to the label value of the feature vector of the sample of the multi-task model and the output of the task-exclusive network.
  • the task exclusive network includes a deep neural network, a convolutional neural network or a self-attention network; the shared network includes a deep neural network, a convolutional neural network or a self-attention network.
  • Fig. 7 is a schematic diagram of a promotional content processing device according to an exemplary disclosed embodiment. As shown in Fig. 7, the promotional content processing device 700 includes:
  • An acquisition module 701, configured to acquire the attribute of the user's behavior on the promotion content
  • the reasoning module 702 is configured to obtain the reasoning result of the multi-task model according to the attribute of the user's behavior on the promotion content and the multi-task model; the multi-task model is generated based on the features corresponding to each task in the related task set The sample of the multi-task model is obtained, the related task set is obtained based on the mutual information between different tasks, and the different tasks are tasks in the initial task set constructed by the features used to construct the task; the reasoning result Including the conversion rate of the promotion content, the playing duration of the promotion content, the presentation type of the promotion content, or the information of the promotion object in the promotion content;
  • the processing module 703 is configured to adjust a promotion strategy for the promotion content according to the reasoning result.
  • FIG. 8 shows a schematic structural diagram of an electronic device 800 suitable for implementing an embodiment of the present disclosure
  • the electronic device is used to realize the functions corresponding to the modeling apparatus 600 of the multi-task model shown in FIG. 6 , Or it is used to implement the functions corresponding to the promotional content processing apparatus 700 shown in FIG. 7 .
  • the electronic device shown in FIG. 8 is only an example, and should not limit the functions and scope of use of the embodiments of the present disclosure.
  • an electronic device 800 may include a processing device (such as a central processing unit, a graphics processing unit, etc.) Various appropriate actions and processes are executed by programs in the memory (RAM) 803 . In the RAM 803, various programs and data necessary for the operation of the electronic device 800 are also stored.
  • the processing device 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804.
  • An input/output (I/O) interface 805 is also connected to the bus 804 .
  • the following devices can be connected to the I/O interface 805: input devices 806 including, for example, a touch screen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; including, for example, a liquid crystal display (LCD), speaker, vibration an output device 807 such as a computer; a storage device 808 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 809.
  • the communication means 809 may allow the electronic device 800 to communicate with other devices wirelessly or by wire to exchange data. While FIG. 8 shows electronic device 800 having various means, it is to be understood that implementing or having all of the means shown is not a requirement. More or fewer means may alternatively be implemented or provided.
  • embodiments of the present disclosure include a computer program product, which includes a computer program carried on a non-transitory computer readable medium, where the computer program includes program code for executing the method shown in the flowchart.
  • the computer program may be downloaded and installed from a network via communication means 809, or from storage means 808, or from ROM 802.
  • the processing device 801 the above-mentioned functions defined in the methods of the embodiments of the present disclosure are executed.
  • the computer-readable medium mentioned above in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two.
  • a computer readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to, electrical connections with one or more wires, portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable Programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.
  • a computer-readable storage medium may be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.
  • a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave carrying computer-readable program code therein. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • a computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device .
  • Program code embodied on a computer readable medium may be transmitted by any appropriate medium, including but not limited to wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.
  • the client and the server can communicate using any currently known or future network protocols such as HTTP (HyperText Transfer Protocol, Hypertext Transfer Protocol), and can communicate with digital data in any form or medium
  • HTTP HyperText Transfer Protocol
  • the communication eg, communication network
  • Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), internetworks (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future developed network of.
  • the above-mentioned computer-readable medium may be included in the above-mentioned electronic device, or may exist independently without being incorporated into the electronic device.
  • the above-mentioned computer-readable medium carries one or more programs, and when the above-mentioned one or more programs are executed by the electronic device, the electronic device:
  • the initial task set includes: the conversion rate of promotional content, the playing duration of the promotional content, and the presentation type of the promotional content and at least two of the information on promotion objects in the promotion content;
  • a related task set is obtained, and the mutual information of the tasks included in the related task set satisfies a first preset condition;
  • the inference result of the multi-task model is obtained;
  • the multi-task model is generated based on the features corresponding to each task in the related task set. samples, the set of related tasks is obtained based on the mutual information between different tasks, and the different tasks are the tasks in the initial task set constructed by the features used to construct the tasks;
  • the reasoning results include the conversion rate of the promotion content , the playback duration of the promotion content, the presentation type of the promotion content, or the information of the promotion object in the promotion content;
  • the promotion strategy for the promotion content is adjusted.
  • Computer program code for carrying out operations of the present disclosure may be written in one or more programming languages, or combinations thereof, including but not limited to object-oriented programming languages—such as Java, Smalltalk, C++, and Includes conventional procedural programming languages - such as "C" or similar programming languages.
  • the program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.
  • the remote computer may be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (for example, using an Internet service provider to connected via the Internet).
  • LAN local area network
  • WAN wide area network
  • Internet service provider for example, using an Internet service provider to connected via the Internet.
  • each block in a flowchart or block diagram may represent a module, program segment, or portion of code that contains one or more logical functions for implementing specified executable instructions.
  • the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations can be implemented by a dedicated hardware-based system that performs the specified functions or operations , or may be implemented by a combination of dedicated hardware and computer instructions.
  • the modules involved in the embodiments described in the present disclosure may be implemented by software or by hardware. Wherein, the name of the module does not constitute a limitation of the module itself under certain circumstances, for example, the first obtaining module may also be described as "a module for obtaining at least two Internet Protocol addresses".
  • FPGAs Field Programmable Gate Arrays
  • ASICs Application Specific Integrated Circuits
  • ASSPs Application Specific Standard Products
  • SOCs System on Chips
  • CPLD Complex Programmable Logical device
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer disks, hard disks, Random Access Memory (RAM), Read Only Memory (ROM), Erasable Programmable Read Only Memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM Random Access Memory
  • ROM Read Only Memory
  • EPROM Erasable Programmable Read Only Memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.
  • Example 1 provides a modeling method of a multi-task model, acquiring features used to construct tasks, and constructing an initial task set according to the features used to construct tasks;
  • the initial The task set includes: at least two of the conversion rate of the promotion content, the playing duration of the promotion content, the presentation type of the promotion content, and the information of the promotion object in the promotion content; determine the different tasks in the initial task set Mutual information between different tasks; according to the mutual information between different tasks, a related task set is obtained, and the mutual information of the tasks included in the related task set satisfies the first preset condition; according to each task in the related task set Generate a sample of the multi-task model corresponding to the feature, and use the sample of the multi-task model to perform model training to obtain the multi-task model.
  • Example 2 provides the method of Example 1, the method further comprising:
  • Example 3 provides the method of Example 2, the aggregating the tasks satisfying the second preset condition in the set of related tasks includes:
  • Example 4 provides the method of Example 3, the aggregated set of related tasks includes multiple task groups, and the method further includes:
  • a plurality of target tasks are determined according to the number of positive samples of the tasks in the task grouping, the sum of the number of positive samples of the multiple target tasks is greater than the threshold of the number of positive samples, and the threshold of the number of positive samples is The minimum value of the sum of the number of positive samples in each task group;
  • a sample of a multi-task model is generated according to features corresponding to the target task in the task group.
  • Example 5 provides the method of Example 4, wherein determining multiple target tasks according to the number of positive samples of tasks in the task grouping includes:
  • the multiple tasks in the task group are multiple target tasks.
  • Example 6 provides the method of Example 1, and performing model training using samples of the multi-task model includes:
  • Example 7 provides the method of Example 6, the task-exclusive network includes a deep neural network, a convolutional neural network, or a self-attention network; the shared network includes a deep neural network, a volume product neural network or self-attention network.
  • Example 8 provides a method for processing promotional content, including: acquiring the attributes of the user's behavior on the promotional content; model to obtain the inference result of the multi-task model; the multi-task model is obtained based on samples of the multi-task model generated by the features corresponding to each task in the related task set, and the related task set is based on the interaction between different tasks
  • the information is obtained, the different tasks are the tasks in the initial task set constructed by the characteristics of the task;
  • the reasoning results include the conversion rate of the promotion content, the playing time of the promotion content, and the presentation type of the promotion content Or multiple types of information about the promotion object in the promotion content; according to the reasoning result, adjust the promotion strategy for the promotion content.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Theoretical Computer Science (AREA)
  • Finance (AREA)
  • Strategic Management (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Game Theory and Decision Science (AREA)
  • Software Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marketing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本公开提供了一种多任务模型的建模方法、推广内容处理方法及相关装置，涉及人工智能技术领域。该多任务模型的建模方法包括：先获取用于构建任务的特征，根据该用于构建任务的特征构建初始任务集合，然后确定该初始任务集合中不同任务之间的互信息，基于该不同任务之间的互信息，获得相关性较强的相关任务集合。相关任务集合中包括的任务的互信息满足第一预设条件。接着，根据相关任务集合中各任务对应的特征，生成多任务模型的样本，以利用多任务模型的样本对多任务模型进行训练。如此，该方法能够获取相关性较强的任务。进而，在进行多任务学习过程中，能够提高多任务模型的学习效率。

Description

多任务模型的建模方法、推广内容处理方法及相关装置
本公开要求于2021年12月21日提交中国国家知识产权局、申请号为202111573367.9、发明名称为“多任务模型的建模方法、推广内容处理方法及相关装置”的中国专利申请的优先权，其全部内容通过引用结合在本公开中。
技术领域
本公开属于人工智能技术领域，具体涉及多任务模型的建模方法、推广内容处理方法、装置、设备、计算机可读存储介质以及计算机程序产品。
背景技术
随着计算机技术尤其是人工智能技术的发展，人工智能技术的应用场景越来越广泛。例如，在推广内容(例如广告)推送的场景中，可以基于人工智能技术预测推广内容的转化率，进而基于该转化率向用户推送该推广内容。
为了提高模型的泛化能力，提高预测的转化率的准确性，通常采用多任务学习(muti task learning，MTL)的方式构建多任务模型。目前，基于人工主观经验选择多个任务，进行多任务学习，以使不同任务之间共享已学到的特征，进而提高多任务模型的学习效率以及泛化能力。
然而，依赖主观经验选择的多个任务会存在相关性较差的情况，会导致多个任务之间存在负面影响的情况，降低多任务模型的学习效率。
发明内容
本公开的目的在于：提供了一种多任务模型的建模方法、推广内容处理方法、装置、设备、计算机可读存储介质以及计算机程序产品，能够提高多任务模型的学习效率。
第一方面，本公开提供了一种多任务模型的建模方法，包括：
获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播 放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
确定所述初始任务集合中不同任务之间的互信息；
根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；
根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
第二方面，本公开提供了一种推广内容处理方法，包括：
获取用户对推广内容的行为的属性；
根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
根据所述推理结果，调整对所述推广内容的推广策略。
第三方面，本公开提供了一种多任务模型的建模装置，包括：
获取模块，用于获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
互信息确定模块，用于确定所述初始任务集合中不同任务之间的互信息；
相关任务确定模块，用于根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；
训练模块，用于根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
第四方面，本公开提供了一种推广内容处理装置，其特征在于，包括：
获取模块，用于获取用户对推广内容的行为的属性；
推理模块，用于根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
处理模块，用于根据所述推理结果，调整对所述推广内容的推广策略。
第五方面，本公开提供一种计算机可读介质，其上存储有计算机程序，该程序被处理装置执行时实现本公开第一方面或第二方面中任一项所述方法的步骤。
第六方面，本公开提供一种电子设备，包括：
存储装置，其上存储有计算机程序；
处理装置，用于执行所述存储装置中的所述计算机程序，以实现本公开第一方面或第二方面中任一项所述方法的步骤。
第七方面，本公开提供了一种包含指令的计算机程序产品，当其在设备上运行时，使得设备执行上述第一方面或第二方面的任一种实现方式所述的方法。
从以上技术方案可以看出，本公开具有如下优点：
本公开提供了多任务模型的建模方法，该方法中，先获取用于构建任务的特征，基于该用于构建任务的特征构建初始任务集合，然后确定该初始任务集合中不同任务之间的互信息，基于该互信息对初始任务集合中进行筛选，得到相关性较强的相关任务集合。接着，基于该相关任务集合中各任务对应的特征，生成多任务模型的样本，以对多任务模型进行训练。与单纯依赖主观经验选择的多个任务相比，本公开对初始任务集合进行筛选后得到的相关任务集合中的任务之间的相关性较强。如此，在进行多任务学习过程中，能够提高多任务模型的学习效率。
本公开的其他特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
附图用来提供对本发明的进一步理解，并且构成说明书的一部分，与本发明实施例一起用于解释本发明，并不构成对本发明的限制。在附图中：
图1为本公开实施例提供的一种多任务模型的建模方法的流程图；
图2为本公开实施例提供的一种获取相关任务集合的示意图；
图3为本公开实施例提供的一种多任务模型的示意图；
图4为本公开实施例提供的又一种多任务模型的建模方法的流程图；
图5为本公开实施例提供的一种推广内容处理方法的流程图；
图6为本公开实施例提供的一种多任务模型的建模装置的示意图；
图7为本公开实施例提供的一种推广内容处理装置的示意图；
图8为本公开实施例提供的一种电子设备的结构示意图。
具体实施方式
本公开实施例中的术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括一个或者更多个该特征。
首先对本公开实施例中所涉及到的一些技术术语进行介绍。
多任务模型指基于多任务学习的方式构建的模型。多任务学习的核心在于：多个任务并行训练，并且互相共享已学到的特征。在推广内容推送场景中，通常也需要对多项内容进行预测，每一项内容可以抽象为一个任务，例如预测转化率、预测潜在用户、预测用户是否点击推广内容等。基于此，可以基于多任务学习的方式，构建多任务模型，以对多项内容进行预测。
在多任务学习过程中，多个任务之间的相关性较强时，多个任务共享已学到的特征，可以提高多任务模型的学习效率。多个任务之间的相关性较弱时，多个任务共享已学到的特征，多个任务之间会出现负迁移现象。即，多个任务之间互相产生负面影响，使每个任务预测的准确率均降低，降低多任务模型的学习效率。
目前，主要依赖主观经验来选择多个任务，并构建多任务模型。基于主观经验选择的多个任务会存在相关性较差的情况，进而降低多任务模型的学习效率。
有鉴于此，本公开实施例提供了一种多任务模型的建模方法，该方法可以由电子设备执行。电子设备可以是服务器。服务器可以是云服务器，例如是中心云计算集群中的中心服务器，或者是边缘云计算集群中的边缘服务器。当然，服务器也可以是本地数据中心中的服务器。本地数据中心是指用户直接控制的数据中心。
该多任务模型的建模方法包括：电子设备获取用于构建任务的特征，根据用于构建任务的特征构建初始任务集合，然后确定初始任务集合中不同任务之间的互信息。基于不同任务之间的互信息，得到相关任务集合，相关任务集合中的互信息满足第一预设条件。根据该相关任务集合中各任务对应的特征，生成多任务模型的样本，利用多任务模型的样本进行模型训练，获得多任务模型。
可见，在本公开实施例提供的多任务模型的建模方法中，电子设备先对初始任务集合进行了筛选，得到的相关任务集合中的任务之间的相关性较强。与单纯地依赖主观经验选择的多个任务相比，提高了多任务模型对应的多个任务之间的相关性。如此，在进行多任务学习过程中，能够提高多任务模型的学习效率。
利用本公开实施例提供的多任务模型的建模方法得到的多任务模型可以应用于多种场景。例如，推广内容推送场景中，多任务模型可以用于预测推广内容的转化率、推广内容的播放时长(推广内容开始播放起直至用户关闭该推广内容所经过的时长)。电子设备可以将用户对推广内容的行为的属性，输入给多任务模型，进而得到上述推理结果。接着电子设备可以基于上述推理结果，调整该推广内容的推广策略。例如推广内容的转化率大于或等于转化率阈值时，且推广内容的播放时长大于或等于时长阈值时，提高该推广内容的推广次数。推广内容的转化率小于转化率阈值时，且推广内容的播放时长小于时长阈值时，降低该推广内容的推广次数。
为了使得本公开的技术方案更加清楚、易于理解，下面以电子设备的角度，对本公开实施例提供的多任务模型的建模方法进行介绍。如图1所 示，该图为本公开实施例提供的一种多任务模型的建模方法的流程图，该方法包括：
S101：电子设备获取用于构建任务的特征，根据该用于构建任务的特征构建初始任务集合。
电子设备可以获取初始数据，该初始数据包括多个特征。举例说明，在推广内容推送场景中，初始数据可以包括推广内容的转化率、推广内容的播放时长、推广内容的呈现类型、推广内容中推广对象的信息、推广内容的生成时间等特征。
电子设备可以基于初始数据获取用于构建任务的特征。在一些示例中，电子设备可以接收开发人员配置的多个特征，作为用于构建任务的特征。例如，电子设备可以通过显示装置(例如显示屏)向用户呈现多个候选特征，如上述初始数据所包括的特征，接着根据开发人员对多个候选特征的选择操作，将开发人员所选择的特征作为用于构建任务的特征。接着，电子设备基于该用于构建任务的特征，构建初始任务集合。
S102：电子设备确定初始任务集合中不同任务之间的互信息。
互信息(mutual information，MI)是指两个随机变量之间的关联程度，即给定一个随机变量后，另一个随机变量不确定性的削弱程度。例如，互信息取值为0(最小值)时，表明给定一个随机变量对确定另一个随机变量没有关系。互信息取值为随机变量的熵(最大值)时，表明给定一个随机变量，能够完全消除另一个随机变量的不确定性。
如图2所示，电子设备可以分别计算初始任务集合中不同任务之间的互信息。例如电子设备可以基于初始任务集合中任务的特征，计算不同任务之间的互信息。具体地，电子设备可以基于如下公式计算初始任务集合中不同任务之间的互信息：
Figure PCTCN2022124765-appb-000001
其中，I(X；Y)为初始任务集合中任务X和任务Y之间的互信息；x为任务X对应的特征，y为任务Y对应的特征；p(x，y)为任务X和任务Y同时发生的概率，p(x)为任务X发生的概率，p(y)为任务Y发生的概率。
需要说明的是，初始任务集合中不同任务之间的互信息是非负的。即，I(X；Y)≥0。并且，I(Y；X)＝I(X；Y)，即，I(Y；X)与I(X；Y)均表示初始任务集合中任务X和任务Y之间的互信息。
S103：电子设备根据不同任务之间的互信息，获得相关任务集合，该相关任务集合中包括的任务的互信息满足第一预设条件。
如上述，互信息可以表征多个随机变量中一个随机变量对另一个随机变量的不确定性。基于此，电子设备可以基于初始任务集合中不同任务之间的互信息，对初始任务集合进行筛选，进而得到相关任务集合。该相关任务集合中包括的任务的互信息满足第一预设条件。例如，第一预设条件可以是不同任务之间的互信息大于或等于互信息阈值。
在一些示例中，电子设备可以接收开发人员配置的互信息阈值，以基于该配置的互信息阈值来确定相关任务集合，从而使得开发人员在一定的范围内，对相关任务集合中的任务进行调整。
本公开实施例不具体限定互信息阈值，如上述，互信息阈值可以为开发人员配置的数值，也可以是默认值。例如互信息阈值可以为0.1、0.2等。
S104：电子设备根据相关任务集合中各任务对应的特征，生成多任务模型的样本，利用该多任务模型的样本进行模型训练，获得该多任务模型。
在确定相关任务集合后，电子设备可以基于相关任务集合中的任务构建多任务模型。由于相关任务集合中的任务之间的互信息满足第一预设条件，表明该相关任务集合中的任务之间的相关性较强。因此，基于相关性较强的任务构建多任务模型后，该多任务模型的学习效率较高。
电子设备可以基于相关任务集合中各任务对应的特征，生成多任务模型的样本。多任务模型的样本包括样本的特征向量及样本的特征向量的标签。如上述，初始数据包括多个特征，相关任务集合中的任务对应的特征为初始数据的多个特征中的部分特征。基于此，电子设备可以将相关任务集合中任务对应的特征作为样本的特征向量的标签，从初始数据中去掉上述相关任务集合中任务对应的特征后，得到的初始数据作为样本的特征向量。
举例说明，初始数据包括特征1、特征2、特征3、特征4和特征5，相关任务集合中任务1对应特征1、任务2对应特征2。电子设备可以将特 征1作为用于监控任务1的样本的特征向量的标签、将特征2作为用于监控任务2的样本的特征向量的标签，从初始数据中去掉特征1和特征2，得到包括特征3、特征4和特征5的初始数据，将该包括特征3、特征4和特征5的初始数据作为多任务模型的样本的特征向量。
接着，电子设备可以利用多任务模型的样本，进行模型训练，得到多任务模型。如图3所示，该图为本公开实施例提供的一种多任务模型的示意图。该多任务模型包括相关任务集合中的多个任务各自对应的任务独占网络310和输出网络330、以及共享网络320。其中，任务独占网络310可以是深度神经网络(deep neural networks，DNN)、卷积神经网络(convolutional neural network，CNN)或自注意力网络；类似的，共享网络320也可以是深度神经网络、卷积神经网络或自注意力网络。
需要说明的是，本公开不具体限定任务独占网络310和共享神经网络320的类型。在图像场景下，为了构建内容信息的不变性，可以利用卷积神经网络构建共享网络320和任务独占网络310。在推广内容场景下，为了满足对特征交叉的需求，可以利用深度神经网络构建共享网络320和任务独占网络310。在文本场景下，为了满足对时序信息并行处理的需求，可以利用自注意力网络构建共享网络320和任务独占网络310。输出网络330可以采用sigmoid、relu等激活函数得到各个任务对应的输出(例如可以是分类值、回归值)。
在一些示例中，电子设备可以将多任务模型的样本的特征向量输入共享网络320，得到共享分量，然后将该共享分量输入给相关任务集合中各个任务的任务独占网络，得到各个任务独占网络的输出。接着电子设备可以基于多任务模型的样本的特征向量的标签值，以及任务独占网络的输出确定损失值，基于该损失值更新任务独占网络的权重，以进行模型训练，进而得到该多任务模型。
基于上述内容描述，本公开实施例提供了一种多任务模型的建模方法，与单纯依赖主观经验选择的多个任务相比，电子设备对初始任务集合进行筛选后得到的相关任务集合中的任务之间的相关性较强。如此，在进行多任务学习过程中，能够提高多任务模型的学习效率。进一步，满足多种场景下的业务需求，实现对多项内容更为准确地预测。
本公开实施例还提供了一种多任务模型的建模方法，该方法在图2所示的实施例的基础上，对相关任务集合中的任务进一步分组聚合，从而平衡聚合相关任务集合中多个任务分组的正样本数量。如图4所示，该多任务模型的建模方法还包括：
S401：电子设备对相关任务集合中满足第二预设条件的任务进行聚合。
第二预设条件可以是相关任务集合中任务相关性大于相关性阈值的多个任务。电子设备可以基于该第二预设条件对相关任务集合中的任务进行聚合，聚合后的相关任务集合包括多个任务分组。
继续参见图3，电子设备可以获取相关任务集合中多个任务对应的任务独占网络310的输出之间的相关性，然后基于该任务独占网络310的输出之间的相关性，对相关任务集合中的多个任务进行聚合。具体地，电子设备可以基于如下公式，计算相关任务集合中多个任务对应的任务独占网络310的输出之间的相关性：
Figure PCTCN2022124765-appb-000002
其中，r mn为相关任务集合中的任务m与任务n之间的相关性；O im为样本i在任务m对应的任务独占网络310的输出，
Figure PCTCN2022124765-appb-000003
为任务m对应的任务独占网络310对所有样本的输出的均值；O in为样本i在任务n对应的任务独占网络310的输出，
Figure PCTCN2022124765-appb-000004
为任务n对应的任务独占网络310对所有样本的输出的均值，K为样本总数。
在一些示例中，电子设备可以将任务m和任务n之间的相关性r mn与相关性阈值进行比较。r mn≥相关性阈值时，将任务m和任务n聚合为一个分组，如此电子设备可以得到聚合后的相关任务集合包括多个任务分组。
S402：电子设备针对每个任务分组，根据任务分组中任务的正样本数量，确定多个目标任务。
其中，多个目标任务的正样本数量之和大于正样本数量阈值，该正样本数量阈值为每个任务分组中正样本数量之和的最小值。举例说明，聚合后的相关任务集合包括多个任务分组。以多个任务分组包括任务分组G1和任务分组G2为例，其中，任务分组G1包括任务T1、任务T2、任务T3， 任务T1的正样本数量为100、任务T2的正样本数量为200、任务T3的正样本数量为300；任务分组G2包括任务T4、任务T5和任务T6，任务T4的正样本数量为200、任务T5的正样本数量为400、任务T6的正样本数量为500。可见，任务分组G1的正样本数量之和为600，任务分组G2的正样本数量之和为1100。电子设备可以将任务分组G1的正样本数量之和600作为正样本数量阈值。
在一些示例中，电子设备可以对任务分组G2中任务T4、任务T5和任务T6进行排序，如按照正样本数量由小到大的顺序进行逐个加和，直至任务分组G2中任务的正样本数量之和大于或等于上述正样本数量阈值600。电子设备可以确定任务分组中任务正样本数量之和大于或等于正样本数量阈值时，任务分组中的多个任务为多个目标任务。如此，电子设备在将任务T4的正样本数量与任务T5的正样本数量进行加和后，加和结果等于正样本数量阈值，电子设备可以将任务T4和任务T5作为该任务分组中的多个目标任务。
在另一些示例中，电子设备可以在任务分组G2随机选择多个任务的正样本数量进行加和，直至多个任务的正样本数量大于或等于正样本数量阈值，接着将随机选择的多个任务确定为多个目标任务。延续上例，电子设备可以将任务分组G2中的任务T5和任务T6的正样本数量进行加和，加和结果大于正样本数量阈值。如此，电子设备可将任务T5和任务T6确定为多个目标任务。
需要说明的是，任务分组G2中任务T4和任务T5的正样本数量之和大于或等于正样本数量阈值是指：当减去任务分组G2中的任务T4或任务T5中任一个的正样本数量时，会使剩余任务的正样本数量之和小于正样本数量阈值。
电子设备确定任务分组的多个目标任务后，可以将任务分组中除该目标任务之外的任务去除。以任务分组G2中的多个目标任务为任务T4和任务T5为例，电子设备可以将任务T6从任务分组G2中去除。
在一些实施例中，电子设备可以基于该任务分组中的多个目标任务对应的特征，生成多任务模型的样本。以任务分组G2包括的目标任务为任务T4和任务T5为例，针对该任务分组G2，将任务T4对应的正样本的特征 向量的标签作为该任务分组G2的标签，以及将任务T5对应的正样本的特征向量的标签也作为该任务分组G2的标签。如此，在本实施例中，电子设备能够将多个任务分组包括的正样本数量进行均衡配置，减少因正样本数量差异较大而使多任务模型的学习效率降低的情况。
基于上述内容描述，多任务模型的建模方法通过对包含多个特征的大规模数据进行统计信息的分析，基于相关性确定相关任务集合，实现对高相关任务的聚合。利用高相关任务进行多任务模型训练，能够减少负向迁移，提高多任务模型的学习效率。
进一步，基于确定的相关任务集合构建多任务模型的结构，结合皮尔逊系数完成带有模型泛化信息的任务聚合。以正样本数量最小的任务分组对应的正样本数量为正样本数量阈值进行各分组任务筛选与组合，从而实现不同任务正样本量的均衡。经过分组聚合后的不同任务间正样本数量可以做到完全等比例，使多任务模型无偏向的完成多任务学习过程。
本公开实施例还提供了一种推广内容处理方法，如图5所示，该推广内容处理方法包括：
S501：电子设备获取用户对推广内容的行为的属性。
用户对推广内容的行为的属性可以包括用户对推广内容进行观看的时长、用户是否对推广内容进行点击操作、用户所点击的推广内容的呈现类型(例如视频类型、图片类型等)、用户是否被推广内容转化等等。
需要说明的是，电子设备需要提前获取用户的授权，在获取到用户对相应数据(如上述用户对推广内容的行为的属性)的授权使用后，电子设备才能获取到用户对推广内容的行为的属性等数据。
S502：电子设备根据用户对推广内容的行为的属性和多任务模型，获得多任务模型的推理结果。
其中，多任务模型基于相关任务集合中各任务对应的特征生成的多任务模型的样本得到。相关任务集合基于不同任务之间的互信息得到，不同任务为用于构建任务的特征所构建的初始任务集合中的任务。推理结果包括推广内容的转化率、推广内容的播放时长、推广内容的呈现类型或推广内容中推广对象的信息中的多种。训练多任务模型的过程可以参见上述实施例中介绍，此处不再赘述。
S503：电子设备根据推理结果，调整对推广内容的推广策略。
在一些示例中，当推理结果表明推广内容的转化率大于转化率阈值，且推广内容的播放时长大于时长阈值时，电子设备提高该推广内容的推广次数。当推广内容的转化率小于转化率阈值时，且推广内容的播放时长小于时长阈值时，电子设备降低该推广内容的推广次数。在另一些示例中，推广内容的呈现类型为预设类型(例如视频类型)，且推广内容中推广对象的信息为预设信息(例如指示该推广对象为游戏、虚拟物品或实物)时，电子设备提高该推广内容的推广次数。如此，减少推广内容的无效投放，减少资源浪费。
图6是根据一示例性公开实施例示出的一种多任务模型的建模装置的示意图，如图6所示，所述多任务模型的建模装置600包括：
获取模块601，用于获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
互信息确定模块602，用于确定所述初始任务集合中不同任务之间的互信息；
相关任务确定模块603，用于根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；
训练模块604，用于根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
可选的，所述相关任务确定模块603，还用于对所述相关任务集合中满足第二预设条件的任务进行聚合。
可选的，所述相关任务确定模块603，具体用于确定所述相关任务集合中多个任务对应的任务独占网络的输出之间的相关性；根据所述相关性对所述相关任务集合中的多个任务进行聚合。
可选的，所述相关任务确定模块603，还用于针对每个任务分组，根据所述任务分组中任务的正样本数量，确定多个目标任务，所述多个目标 任务的正样本数量之和大于正样本数量阈值，所述正样本数量阈值为每个任务分组中正样本数量之和的最小值；
所述训练模块604，具体用于根据所述任务分组中所述目标任务对应的特征，生成多任务模型的样本。
可选的，所述相关任务确定模块603，具体用于按照所述任务分组中任务的正样本数量由小到大的顺序进行逐个加和，直至所述任务分组中任务的正样本数量之和大于或等于所述正样本数量阈值；确定所述任务分组中任务正样本数量之和大于或等于所述正样本数量阈值时，所述任务分组中的多个任务为多个目标任务。
可选的，所述训练模块604，具体用于将所述多任务模型的样本的特征向量输入共享网络，得到共享分量；将所述共享分量输入给所述相关任务集合中各个任务的任务独占网络，得到所述各个任务的任务独占网络对应的输出；根据所述多任务模型的样本的特征向量的标签值，以及所述任务独占网络的输出，训练所述多任务模型。
可选的，所述任务独占网络包括深度神经网络、卷积神经网络或自注意力网络；所述共享网络包括深度神经网络、卷积神经网络或自注意力网络。
图7是根据一示例性公开实施例示出的一种推广内容处理装置的示意图，如图7所示，所述推广内容处理装置700包括：
获取模块701，用于获取用户对推广内容的行为的属性；
推理模块702，用于根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
处理模块703，用于根据所述推理结果，调整对所述推广内容的推广策略。
上述各模块的功能在上一实施例中的方法步骤中已详细阐述，在此不 做赘述。
下面参考图8，其示出了适于用来实现本公开实施例的电子设备800的结构示意图，该电子设备用于实现如图6所示的多任务模型的建模装置600对应的功能，或用于实现如图7所示的推广内容处理装置700对应的功能。图8示出的电子设备仅仅是一个示例，不应对本公开实施例的功能和使用范围带来任何限制。
如图8所示，电子设备800可以包括处理装置(例如中央处理器、图形处理器等)801，其可以根据存储在只读存储器(ROM)802中的程序或者从存储装置808加载到随机访问存储器(RAM)803中的程序而执行各种适当的动作和处理。在RAM 803中，还存储有电子设备800操作所需的各种程序和数据。处理装置801、ROM 802以及RAM 803通过总线804彼此相连。输入/输出(I/O)接口805也连接至总线804。
通常，以下装置可以连接至I/O接口805：包括例如触摸屏、触摸板、键盘、鼠标、摄像头、麦克风、加速度计、陀螺仪等的输入装置806；包括例如液晶显示器(LCD)、扬声器、振动器等的输出装置807；包括例如磁带、硬盘等的存储装置808；以及通信装置809。通信装置809可以允许电子设备800与其他设备进行无线或有线通信以交换数据。虽然图8示出了具有各种装置的电子设备800，但是应理解的是，并不要求实施或具备所有示出的装置。可以替代地实施或具备更多或更少的装置。
特别地，根据本公开的实施例，上文参考流程图描述的过程可以被实现为计算机软件程序。例如，本公开的实施例包括一种计算机程序产品，其包括承载在非暂态计算机可读介质上的计算机程序，该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实施例中，该计算机程序可以通过通信装置809从网络上被下载和安装，或者从存储装置808被安装，或者从ROM 802被安装。在该计算机程序被处理装置801执行时，执行本公开实施例的方法中限定的上述功能。
需要说明的是，本公开上述的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件，或者任意以上的组合。计算机可读存储介质的更 具体的例子可以包括但不限于：具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本公开中，计算机可读存储介质可以是任何包含或存储程序的有形介质，该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本公开中，计算机可读信号介质可以包括在基带中或者作为载波一部分传播的数据信号，其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式，包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读信号介质还可以是计算机可读存储介质以外的任何计算机可读介质，该计算机可读信号介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输，包括但不限于：电线、光缆、RF(射频)等等，或者上述的任意合适的组合。
在一些实施方式中，客户端、服务器可以利用诸如HTTP(HyperText Transfer Protocol，超文本传输协议)之类的任何当前已知或未来研发的网络协议进行通信，并且可以与任意形式或介质的数字数据通信(例如，通信网络)互连。通信网络的示例包括局域网(“LAN”)，广域网(“WAN”)，网际网(例如，互联网)以及端对端网络(例如，ad hoc端对端网络)，以及任何当前已知或未来研发的网络。
上述计算机可读介质可以是上述电子设备中所包含的；也可以是单独存在，而未装配入该电子设备中。
上述计算机可读介质承载有一个或者多个程序，当上述一个或者多个程序被该电子设备执行时，使得该电子设备：
获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
确定所述初始任务集合中不同任务之间的互信息；
根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务 集合中包括的任务的互信息满足第一预设条件；
根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型；或者，
获取用户对推广内容的行为的属性；
根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
根据所述推理结果，调整对所述推广内容的推广策略。
可以以一种或多种程序设计语言或其组合来编写用于执行本公开的操作的计算机程序代码，上述程序设计语言包括但不限于面向对象的程序设计语言—诸如Java、Smalltalk、C++，还包括常规的过程式程序设计语言——诸如“C”语言或类似的程序设计语言。程序代码可以完全地在用户计算机上执行、部分地在用户计算机上执行、作为一个独立的软件包执行、部分在用户计算机上部分在远程计算机上执行、或者完全在远程计算机或服务器上执行。在涉及远程计算机的情形中，远程计算机可以通过任意种类的网络——包括局域网(LAN)或广域网(WAN)——连接到用户计算机，或者，可以连接到外部计算机(例如利用因特网服务提供商来通过因特网连接)。
附图中的流程图和框图，图示了按照本公开各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分，该模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以 及框图和/或流程图中的方框的组合，可以用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
描述于本公开实施例中所涉及到的模块可以通过软件的方式实现，也可以通过硬件的方式来实现。其中，模块的名称在某种情况下并不构成对该模块本身的限定，例如，第一获取模块还可以被描述为“获取至少两个网际协议地址的模块”。
本文中以上描述的功能可以至少部分地由一个或多个硬件逻辑部件来执行。例如，非限制性地，可以使用的示范类型的硬件逻辑部件包括：现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、片上系统(SOC)、复杂可编程逻辑设备(CPLD)等等。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
根据本公开的一个或多个实施例，示例1提供了一种多任务模型的建模方法，获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；确定所述初始任务集合中不同任务之间的互信息；根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
根据本公开的一个或多个实施例，示例2提供了示例1的方法，所述方法还包括：
对所述相关任务集合中满足第二预设条件的任务进行聚合。
根据本公开的一个或多个实施例，示例3提供了示例2的方法，所述对所述相关任务集合中满足第二预设条件的任务进行聚合，包括：
确定所述相关任务集合中多个任务对应的任务独占网络的输出之间的相关性；
根据所述相关性对所述相关任务集合中的多个任务进行聚合。
根据本公开的一个或多个实施例，示例4提供了示例3的方法，聚合后的所述相关任务集合包括多个任务分组，所述方法还包括：
针对每个任务分组，根据所述任务分组中任务的正样本数量，确定多个目标任务，所述多个目标任务的正样本数量之和大于正样本数量阈值，所述正样本数量阈值为每个任务分组中正样本数量之和的最小值；
所述根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，包括：
根据所述任务分组中所述目标任务对应的特征，生成多任务模型的样本。
根据本公开的一个或多个实施例，示例5提供了示例4的方法，所述根据所述任务分组中任务的正样本数量，确定多个目标任务，包括：
按照所述任务分组中任务的正样本数量由小到大的顺序进行逐个加和，直至所述任务分组中任务的正样本数量之和大于或等于所述正样本数量阈值；
确定所述任务分组中任务正样本数量之和大于或等于所述正样本数量阈值时，所述任务分组中的多个任务为多个目标任务。
根据本公开的一个或多个实施例，示例6提供了示例1的方法，所述利用所述多任务模型的样本进行模型训练，包括：
将所述多任务模型的样本的特征向量输入共享网络，得到共享分量；
将所述共享分量输入给所述相关任务集合中各个任务的任务独占网络，得到所述各个任务的任务独占网络对应的输出；
根据所述多任务模型的样本的特征向量的标签值，以及所述任务独占 网络的输出，训练所述多任务模型。
根据本公开的一个或多个实施例，示例7提供了示例6的方法，所述任务独占网络包括深度神经网络、卷积神经网络或自注意力网络；所述共享网络包括深度神经网络、卷积神经网络或自注意力网络。
根据本公开的一个或多个实施例，示例8提供了一种推广内容处理方法，包括：获取用户对推广内容的行为的属性；根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；根据所述推理结果，调整对所述推广内容的推广策略。
以上描述仅为本公开的较佳实施例以及对所运用技术原理的说明。本领域技术人员应当理解，本公开中所涉及的公开范围，并不限于上述技术特征的特定组合而成的技术方案，同时也应涵盖在不脱离上述公开构思的情况下，由上述技术特征或其等同特征进行任意组合而形成的其它技术方案。例如上述特征与本公开中公开的(但不限于)具有类似功能的技术特征进行互相替换而形成的技术方案。
此外，虽然采用特定次序描绘了各操作，但是这不应当理解为要求这些操作以所示出的特定次序或以顺序次序执行来执行。在一定环境下，多任务和并行处理可能是有利的。同样地，虽然在上面论述中包含了若干具体实现细节，但是这些不应当被解释为对本公开的范围的限制。在单独的实施例的上下文中描述的某些特征还可以组合地实现在单个实施例中。相反地，在单个实施例的上下文中描述的各种特征也可以单独地或以任何合适的子组合的方式实现在多个实施例中。
尽管已经采用特定于结构特征和/或方法逻辑动作的语言描述了本主题，但是应当理解所附权利要求书中所限定的主题未必局限于上面描述的特定特征或动作。相反，上面所描述的特定特征和动作仅仅是实现权利要求书的示例形式。关于上述实施例中的装置，其中各个模块执行操作的具 体方式已经在有关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。

Claims (13)

  1. 一种多任务模型的建模方法，其特征在于，包括：
    获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
    确定所述初始任务集合中不同任务之间的互信息；
    根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；
    根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    对所述相关任务集合中满足第二预设条件的任务进行聚合。
  3. 根据权利要求2所述的方法，其特征在于，所述对所述相关任务集合中满足第二预设条件的任务进行聚合，包括：
    确定所述相关任务集合中多个任务对应的任务独占网络的输出之间的相关性；
    根据所述相关性对所述相关任务集合中的多个任务进行聚合。
  4. 根据权利要求3所述的方法，其特征在于，聚合后的所述相关任务集合包括多个任务分组，所述方法还包括：
    针对每个任务分组，根据所述任务分组中任务的正样本数量，确定多个目标任务，所述多个目标任务的正样本数量之和大于正样本数量阈值，所述正样本数量阈值为每个任务分组中正样本数量之和的最小值；
    所述根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，包括：
    根据所述任务分组中所述目标任务对应的特征，生成多任务模型的样本。
  5. 根据权利要求4所述的方法，其特征在于，所述根据所述任务分组中任务的正样本数量，确定多个目标任务，包括：
    按照所述任务分组中任务的正样本数量由小到大的顺序进行逐个加和， 直至所述任务分组中任务的正样本数量之和大于或等于正样本数量阈值；
    确定所述任务分组中任务正样本数量之和大于或等于所述正样本数量阈值时，所述任务分组中的多个任务为多个目标任务。
  6. 根据权利要求1所述的方法，其特征在于，所述利用所述多任务模型的样本进行模型训练，包括：
    将所述多任务模型的样本的特征向量输入共享网络，得到共享分量；
    将所述共享分量输入给所述相关任务集合中各个任务的任务独占网络，得到所述各个任务的任务独占网络对应的输出；
    根据所述多任务模型的样本的特征向量的标签值，以及所述任务独占网络的输出，训练所述多任务模型。
  7. 根据权利要求6所述的方法，其特征在于，所述任务独占网络包括深度神经网络、卷积神经网络或自注意力网络；所述共享网络包括深度神经网络、卷积神经网络或自注意力网络。
  8. 一种推广内容处理方法，其特征在于，包括：
    获取用户对推广内容的行为的属性；
    根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
    根据所述推理结果，调整对所述推广内容的推广策略。
  9. 一种多任务模型的建模装置，其特征在于，包括：
    获取模块，用于获取用于构建任务的特征，根据所述用于构建任务的特征构建初始任务集合；所述初始任务集合包括：推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型和所述推广内容中推广对象的信息中的至少两种；
    互信息确定模块，用于确定所述初始任务集合中不同任务之间的互信息；
    相关任务确定模块，用于根据所述不同任务之间的互信息，获得相关任务集合，所述相关任务集合中包括的任务的互信息满足第一预设条件；
    训练模块，用于根据所述相关任务集合中各任务对应的特征，生成多任务模型的样本，利用所述多任务模型的样本进行模型训练，获得所述多任务模型。
  10. 一种推广内容处理装置，其特征在于，包括：
    获取模块，用于获取用户对推广内容的行为的属性；
    推理模块，用于根据所述用户对所述推广内容的行为的属性和多任务模型，获得所述多任务模型的推理结果；所述多任务模型基于相关任务集合中各任务对应的特征生成的所述多任务模型的样本得到，所述相关任务集合基于不同任务之间的互信息得到，所述不同任务为用于构建任务的特征所构建的初始任务集合中的任务；所述推理结果包括推广内容的转化率、所述推广内容的播放时长、所述推广内容的呈现类型或所述推广内容中推广对象的信息中的多种；
    处理模块，用于根据所述推理结果，调整对所述推广内容的推广策略。
  11. 一种电子设备，其特征在于，包括：
    存储装置，其上存储有计算机程序；
    处理装置，用于执行所述存储装置中的所述计算机程序，以实现权利要求1至7中任一项所述方法的步骤；或，以实现权利要求8所述方法的步骤。
  12. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理装置执行时实现权利要求1至7中任一项所述方法的步骤；或，该程序被处理装置执行时实现权利要求8所述方法的步骤。
  13. 一种计算机程序产品，其特征在于，当所述计算机程序产品在计算机上运行时，使得计算机执行如权利要求1至7中任一项所述方法的步骤；或，当所述计算机程序产品在计算机上运行时，使得计算机执行如权利要求8所述方法的步骤。
PCT/CN2022/124765 2021-12-21 2022-10-12 多任务模型的建模方法、推广内容处理方法及相关装置 Ceased WO2023116138A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111573367.9 2021-12-21
CN202111573367.9A CN114240506B (zh) 2021-12-21 2021-12-21 多任务模型的建模方法、推广内容处理方法及相关装置

Publications (1)

Publication Number Publication Date
WO2023116138A1 true WO2023116138A1 (zh) 2023-06-29

Family

ID=80760649

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/124765 Ceased WO2023116138A1 (zh) 2021-12-21 2022-10-12 多任务模型的建模方法、推广内容处理方法及相关装置

Country Status (2)

Country Link
CN (1) CN114240506B (zh)
WO (1) WO2023116138A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116879409A (zh) * 2023-07-13 2023-10-13 华南农业大学 一种基于气体传感器的水果果实损伤的分析检测方法
CN116992134A (zh) * 2023-07-07 2023-11-03 中国平安人寿保险股份有限公司 基于多任务模型的信息预测方法、装置、电子设备及介质

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114240506B (zh) * 2021-12-21 2025-04-11 北京有竹居网络技术有限公司 多任务模型的建模方法、推广内容处理方法及相关装置
CN114860411B (zh) * 2022-05-17 2023-05-05 北京百度网讯科技有限公司 多任务学习方法、装置、电子设备和存储介质

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109977402A (zh) * 2019-03-11 2019-07-05 北京明略软件系统有限公司 一种命名实体识别方法及系统
CN110378229A (zh) * 2019-06-19 2019-10-25 浙江大学 一种基于filter–wrapper框架的电子鼻数据特征选择方法
CN111178981A (zh) * 2020-01-02 2020-05-19 众安在线财产保险股份有限公司 一种广告投放方法、装置、计算机设备及存储介质
CN112559007A (zh) * 2020-12-14 2021-03-26 北京百度网讯科技有限公司 多任务模型的参数更新方法、装置及电子设备
CN112561077A (zh) * 2020-12-14 2021-03-26 北京百度网讯科技有限公司 多任务模型的训练方法、装置及电子设备
US20210133807A1 (en) * 2019-10-30 2021-05-06 Target Brands, Inc. Multitask transfer learning for optimization of targeted promotional programs
CN114240506A (zh) * 2021-12-21 2022-03-25 北京有竹居网络技术有限公司 多任务模型的建模方法、推广内容处理方法及相关装置

Family Cites Families (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070218998A1 (en) * 2005-09-12 2007-09-20 Arbogast Christopher P Download and configuration method for gaming machines
US8924315B2 (en) * 2011-12-13 2014-12-30 Xerox Corporation Multi-task learning using bayesian model with enforced sparsity and leveraging of task correlations
US9378065B2 (en) * 2013-03-15 2016-06-28 Advanced Elemental Technologies, Inc. Purposeful computing
CN107045506A (zh) * 2016-02-05 2017-08-15 阿里巴巴集团控股有限公司 评估指标获取方法及装置
CN109165997A (zh) * 2018-07-19 2019-01-08 阿里巴巴集团控股有限公司 一种线下购物推荐内容的生成方法及装置
CN109784476B (zh) * 2019-01-12 2022-08-16 福州大学 一种改进dsod网络的方法
CN111951249A (zh) * 2020-08-13 2020-11-17 浙江理工大学 基于多任务学习网络的手机导光板缺陷可视化检测方法
CN112202901A (zh) * 2020-09-30 2021-01-08 中国银行股份有限公司 一种推广信息的推送方法、装置、设备及介质
CN112163676B (zh) * 2020-10-13 2024-04-05 北京百度网讯科技有限公司 多任务服务预测模型训练方法、装置、设备以及存储介质
CN112967086B (zh) * 2021-02-27 2022-08-02 北京橙色风暴数字技术有限公司 一种智能营销推广方法、装置和电子设备
CN113205362A (zh) * 2021-04-30 2021-08-03 北京有竹居网络技术有限公司 确定推广方的方法、装置、设备、存储介质和程序产品
CN113344086B (zh) * 2021-06-16 2022-07-01 深圳市商汤科技有限公司 人机回圈方法、装置、系统、电子设备和存储介质

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109977402A (zh) * 2019-03-11 2019-07-05 北京明略软件系统有限公司 一种命名实体识别方法及系统
CN110378229A (zh) * 2019-06-19 2019-10-25 浙江大学 一种基于filter–wrapper框架的电子鼻数据特征选择方法
US20210133807A1 (en) * 2019-10-30 2021-05-06 Target Brands, Inc. Multitask transfer learning for optimization of targeted promotional programs
CN111178981A (zh) * 2020-01-02 2020-05-19 众安在线财产保险股份有限公司 一种广告投放方法、装置、计算机设备及存储介质
CN112559007A (zh) * 2020-12-14 2021-03-26 北京百度网讯科技有限公司 多任务模型的参数更新方法、装置及电子设备
CN112561077A (zh) * 2020-12-14 2021-03-26 北京百度网讯科技有限公司 多任务模型的训练方法、装置及电子设备
CN114240506A (zh) * 2021-12-21 2022-03-25 北京有竹居网络技术有限公司 多任务模型的建模方法、推广内容处理方法及相关装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116992134A (zh) * 2023-07-07 2023-11-03 中国平安人寿保险股份有限公司 基于多任务模型的信息预测方法、装置、电子设备及介质
CN116879409A (zh) * 2023-07-13 2023-10-13 华南农业大学 一种基于气体传感器的水果果实损伤的分析检测方法

Also Published As

Publication number Publication date
CN114240506A (zh) 2022-03-25
CN114240506B (zh) 2025-04-11

Similar Documents

Publication Publication Date Title
WO2023116138A1 (zh) 多任务模型的建模方法、推广内容处理方法及相关装置
US20170316347A1 (en) Crowdsourcing System with Community Learning
CN109829433A (zh) 人脸图像识别方法、装置、电子设备及存储介质
CN112800276B (zh) 视频封面确定方法、装置、介质及设备
CN111368973B (zh) 用于训练超网络的方法和装置
CN113222050B (zh) 图像分类方法、装置、可读介质及电子设备
CN115841366A (zh) 物品推荐模型训练方法、装置、电子设备及存储介质
CN118736343A (zh) 目标检测模型的训练方法、目标检测方法、装置及设备
CN116483891A (zh) 一种信息预测方法、装置、设备和存储介质
CN112418233B (zh) 图像处理方法、装置、可读介质及电子设备
CN112861935A (zh) 模型生成方法、对象分类方法、装置、电子设备及介质
WO2024031999A1 (zh) 图像分层方法、装置、电子设备及存储介质
WO2023045870A1 (zh) 网络模型的压缩方法、图像生成方法、装置、设备及介质
CN114201674A (zh) 模型训练方法、推广内容的处理方法及相关装置
CN114611610A (zh) 一种视频流行度预测方法、装置、设备和介质
WO2025020569A1 (zh) 图像数据集处理方法、装置、设备及存储介质
CN113435528B (zh) 对象分类的方法、装置、可读介质和电子设备
CN117201836A (zh) 一种视频画质检测方法、装置、设备和存储介质
CN113837808B (zh) 一种推广信息的推送方法、装置、设备、介质及产品
CN115619456A (zh) 对象推荐方法、装置、介质以及电子设备
CN116010867A (zh) 负样本确定方法、装置、介质及电子设备
CN115688042A (zh) 模型融合方法、装置、设备及存储介质
CN117077036B (zh) 一种模型稳定性检测方法、装置及设备
US12406030B1 (en) Managing resources using a graph inference model to infer tags
US20260030510A1 (en) Image processing method and apparatus, electronic device and storage medium

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22909456

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22909456

Country of ref document: EP

Kind code of ref document: A1