Multi-task model training method, promotion content processing method and related devices

Abstract

The present disclosure provides a multi-task model training method, a promotion content processing method and a related device. The multi-task model includes an auxiliary network, a core network and a forward-looking gradient network. The method includes: obtaining a sample feature vector, the sample feature vector is used to characterize the user's behavior on the promotion content; inputting the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector; inputting the auxiliary feature vector into the forward-looking gradient network to obtain a gradient feature vector, the gradient feature vector characterizes the contribution of the auxiliary feature vector to the optimization target of the core network; after fusing the gradient feature vector with the auxiliary feature vector, inputting the auxiliary output network to obtain the output of the auxiliary output network; updating the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network, until the auxiliary network meets the first training stop condition. The auxiliary task in the multi-task model can be biased to optimize the core task.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence technology, and in particular to a multi-task model training method, a promotion content processing method, and related devices. Background Art

[0002] With the development of computer technology, especially artificial intelligence technology, the application scenarios of artificial intelligence technology are becoming more and more extensive. For example, in the scenario of pushing promotional content (such as advertisements), 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.

[0003] In order to improve the generalization ability of the model and the accuracy of the predicted conversion rate, multi-task learning (MTL) is usually used to build a multi-task model. At present, multiple tasks (conversion rate of promotional content, viewing time of promotional content) are selected based on artificial subjective experience, and multi-task learning is performed to improve the learning efficiency and generalization ability of the multi-task model.

[0004] However, for promotional content, multi-task models usually optimize a core task, such as optimizing the core task of predicting conversion rate. How to make the auxiliary tasks in the multi-task model biasedly optimize the core task in a multi-task scenario is a technical problem that the industry urgently needs to solve. Summary of the invention

[0005] The purpose of the present disclosure is to provide a multi-task model training method, a promotion content processing method, related devices, electronic devices, computer-readable storage media and computer program products, which can enable the auxiliary tasks in the multi-task model to biasly optimize the core tasks in a multi-task scenario to meet business needs.

[0006] In a first aspect, the present disclosure provides a multi-task model training method, wherein the multi-task model includes an auxiliary network, a core network, and a look-ahead gradient network, wherein the auxiliary network includes an auxiliary exclusive network and an auxiliary output network; the method includes:

[0007] Obtaining a sample feature vector, where the sample feature vector is used to characterize the user's behavior towards the promotion content;

[0008] Inputting the sample feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector;

[0009] Inputting the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network;

[0010] After fusing the gradient feature vector and the auxiliary feature vector, the resultant is input into the auxiliary output network to obtain the output of the auxiliary output network;

[0011] The weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

[0012] In a second aspect, the present disclosure provides a promotion content processing method, comprising:

[0013] Obtaining a user feature vector for characterizing the user's behavior towards the promoted content;

[0014] According to the user feature vector and the multi-task model, an inference result of the multi-task model is obtained; the multi-task model is obtained by the method described in the first aspect above;

[0015] According to the inference result, the promotion strategy for the promotion content is adjusted.

[0016] In a third aspect, the present disclosure provides a multi-task model training device, wherein the multi-task model includes an auxiliary network, a core network, and a look-ahead gradient network, wherein the auxiliary network includes an auxiliary exclusive network and an auxiliary output network; the device includes:

[0017] A sample acquisition module, used to acquire a sample feature vector, wherein the sample feature vector is used to characterize the user's behavior towards the promotion content;

[0018] A vector processing module is used to input the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector; input the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network; after fusing the gradient feature vector with the auxiliary feature vector, input the result into the auxiliary output network to obtain the output of the auxiliary output network;

[0019] A model training module is used to update the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

[0020] In a fourth aspect, the present disclosure provides a promotion content processing device, including:

[0021] An acquisition module, used to acquire a user feature vector for characterizing the user's behavior towards the promoted content;

[0022] An inference module, used for obtaining an inference result of the multi-task model according to the user feature vector and the multi-task model; the multi-task model is obtained by the method described in the first aspect above;

[0023] A processing module is used to adjust the promotion strategy for the promotion content according to the inference result.

[0024] In a fifth aspect, the present disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processing device, implements the steps of the method described in any one of the first aspect or the second aspect of the present disclosure.

[0025] In a sixth aspect, the present disclosure provides an electronic device, including:

[0026] a storage device having a computer program stored thereon;

[0027] A processing device is used 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.

[0028] In a seventh aspect, the present disclosure provides a computer program product comprising instructions, which, when executed on a device, enables the device to execute the method described in any one of the implementations of the first or second aspect.

[0029] It can be seen from the above technical solutions that the present disclosure has the following advantages:

[0030] The present disclosure provides a multi-task model training method, the multi-task model includes an auxiliary network, a core network and a forward-looking gradient network, the auxiliary network includes an auxiliary exclusive network and an auxiliary output network, the method includes: first obtaining a sample feature vector, the sample feature vector is used to characterize the user's behavior on the promotion content; then the sample feature vector is input into the auxiliary exclusive network to obtain an auxiliary feature vector, the auxiliary feature vector is input into the forward-looking gradient network to obtain a gradient feature vector, the gradient feature vector characterizes the contribution of the auxiliary feature vector to the optimization target of the core network; then the gradient feature vector is fused with the auxiliary feature vector and input into the auxiliary output network. Since the input of the auxiliary output network includes the contribution to the optimization target of the core network, the auxiliary network can also optimize the optimization target of the core network, so that the auxiliary network can biasedly optimize the core target of the core task. Finally, the weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization targets of the auxiliary network and the core network. If the output of the auxiliary output network satisfies any one of the optimization targets of the auxiliary network and the core network, the sample feature vector is considered to be the feature vector corresponding to the positive sample, so that the auxiliary network can further optimize the optimization target of the core network in a biased manner to meet business needs.

[0031] Other features and advantages of the present disclosure will be described in detail in the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical method of the embodiments of the present disclosure, the drawings required for use in the embodiments are briefly introduced below.

[0033] Figure 1 A flowchart of a multi-task model training method provided in an embodiment of the present disclosure;

[0034] Figure 2 A schematic diagram of a multi-task model provided in an embodiment of the present disclosure;

[0035] Figure 3 A flowchart of a promotion content processing method provided by an embodiment of the present disclosure;

[0036] Figure 4 A schematic diagram of a multi-task model training device provided in an embodiment of the present disclosure;

[0037] Figure 5 A schematic diagram of a promotion content processing device provided by an embodiment of the present disclosure;

[0038] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0039] The terms "first" and "second" in the embodiments of the present disclosure are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0040] First, some technical terms involved in the embodiments of the present disclosure are introduced.

[0041] A multi-task model refers to a model built based on multi-task learning. The core of multi-task learning is that multiple tasks are trained in parallel and the learned features are shared with each other. In the promotional content push scenario, it is usually necessary to predict multiple contents. Each content can be abstracted as a task, such as predicting conversion rate, predicting potential users, predicting whether users click on promotional content, etc. Based on this, a multi-task model can be built based on multi-task learning to predict multiple contents.

[0042] Promotional content refers to content used to promote the promoted object, such as advertisements. Generally, an artificial intelligence model can be pre-trained to predict the conversion rate of users converted by the promoted content, and then the promotional content is subsequently processed based on the conversion rate. For example, when the conversion rate is high, the number of times the promotional content is pushed to the user can be increased, and when the conversion rate is low, the number of times the promotional content is pushed to the user can be reduced. In this way, the ineffective delivery of promotional content can be reduced, and the waste of resources (such as computing resources and storage resources) can be reduced.

[0043] For promotional content, a multi-task model usually only needs to optimize the core task of predicting conversion rate, which can be the optimization target of the core network. How to make the auxiliary tasks in the multi-task model biasedly optimize the core tasks in a multi-task scenario is a technical problem that the industry urgently needs to solve.

[0044] In view of this, an embodiment of the present disclosure provides a multi-task model training method, which can be performed by an electronic device. The electronic device can be a server. The server can be a cloud server, such as a central server in a central cloud computing cluster, or an edge server in an edge cloud computing cluster. Of course, the server can also be a server in a local data center. A local data center refers to a data center directly controlled by a user.

[0045] The multi-task model includes an auxiliary network, a core network and a forward-looking gradient network, wherein the auxiliary network includes an auxiliary exclusive network and an auxiliary output network. The method includes obtaining a sample feature vector, wherein the sample feature vector is used to characterize the user's behavior on the promotion content; inputting the sample feature vector into the forward-looking gradient network to obtain a gradient feature vector, wherein the gradient feature vector characterizes the contribution of the auxiliary feature vector to the optimization target of the core network; then fusing the gradient feature vector with the auxiliary feature vector and inputting it into the auxiliary output network. Since the input of the auxiliary output network includes the contribution to the optimization target of the core network, the auxiliary network can also optimize the optimization target of the core network, thereby enabling the auxiliary network to biasedly optimize the core target of the core task. Finally, the weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network. If the output of the auxiliary output network satisfies any one of the optimization targets of the auxiliary network and the core network, the sample feature vector is considered to be the feature vector corresponding to the positive sample, thereby enabling the auxiliary network to biasedly optimize the optimization target of the core network and meet business needs.

[0046] The multi-task model obtained by the multi-task model training method provided by the embodiment of the present disclosure can be applied to a variety of scenarios. For example, in the promotion content push scenario, the multi-task model can be used to predict the conversion rate of the promotion content and the playback time of the promotion content (the time from the start of the promotion content until the user closes the promotion content). The electronic device can input the user feature vector used to characterize the user's behavior on the promotion content into the multi-task model, and then obtain the above-mentioned reasoning result. The electronic device can then adjust the promotion strategy of the promotion content based on the reasoning result. For example, when the conversion rate of the promotion content is greater than or equal to the conversion rate threshold, and the playback time of the promotion content is greater than or equal to the duration threshold, the number of promotions of the promotion content is increased; when the conversion rate of the promotion content is less than the conversion rate threshold, and the playback time of the promotion content is less than the duration threshold, the number of promotions of the promotion content is reduced, etc.

[0047] In order to make the technical solution of the present disclosure clearer and easier to understand, the multi-task model training method provided by the embodiment of the present disclosure is introduced from the perspective of electronic equipment. Figure 1 As shown in the figure, this figure is a flowchart of a multi-task model training method provided by an embodiment of the present disclosure, the method comprising:

[0048] S101. The electronic device obtains a sample feature vector.

[0049] The sample feature vector is used to characterize the user's behavior towards the promoted content, such as the user's behavior towards the promoted content in a historical period. The user's behavior towards the promoted content may include the length of time the user has watched the promoted content, whether the user has clicked on the promoted content, the presentation type of the promoted content clicked by the user (such as video type, picture type, etc.), whether the user has been converted by the promoted content, etc.

[0050] It should be noted that the electronic device needs to obtain the user's authorization in advance. Only after obtaining the user's authorization to use the corresponding data (such as the user's behavior towards the promotional content mentioned above) can the electronic device obtain data such as the user's behavior towards the promotional content.

[0051] S102: The electronic device inputs the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector.

[0052] For ease of understanding, Figure 2 A schematic diagram of a multi-task model is provided, wherein the multi-task model 200 includes: a shared network 210, an auxiliary network (including an auxiliary exclusive network 220 and an auxiliary output network 230), a look-ahead gradient network 240, and a core network (including a core exclusive network 250 and a core output network 260).

[0053] It should be noted that the above-mentioned shared network 210 can be a deep neural network, a convolutional neural network or a self-attention network, etc.; the above-mentioned auxiliary exclusive network 220 can be a deep neural network, a convolutional neural network or a self-attention network, etc.; the above-mentioned core exclusive network 250 can be a deep neural network, a convolutional neural network or a self-attention network, etc.

[0054] In some examples, the electronic device may input the sample feature vector into the shared network 210 to obtain a shared feature vector output by the shared network 210, and then input the shared feature vector into the auxiliary exclusive network 220 to obtain an auxiliary feature vector.

[0055] S103: The electronic device inputs the auxiliary feature vector into a forward-looking gradient network to obtain a gradient feature vector.

[0056] The gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network. That is to say, through the gradient feature vector, it can be known which dimensional vector in the auxiliary feature vector has a greater contribution to the optimization target of the core network, and which dimensional vector in the auxiliary feature vector has a smaller contribution to the optimization target of the core network. In this way, the auxiliary network can be biased to pay attention to the vector with greater contribution, thereby achieving biased optimization of the optimization target of the core network.

[0057] In some examples, the look-ahead gradient network is updated based on the optimization objective of the core network. For example, the weight of the look-ahead gradient network can be updated based only on the optimization objective of the core network. In some embodiments, the look-ahead loss value of the look-ahead gradient network can be calculated based on the optimization objective of the core network, the output of the look-ahead gradient network, and the objective function of the look-ahead gradient network, and then the weight of the look-ahead gradient network is updated using the look-ahead loss value. The gradient of the feature input to the look-ahead gradient network by the auxiliary exclusive network can be calculated based on the look-ahead loss value as a supervision signal of the look-ahead gradient network, and the mean square error supervision of the gradient feature vector output by the look-ahead gradient network is performed.

[0058] S104: The electronic device fuses the gradient feature vector and the auxiliary feature vector, and inputs the resultant feature vector into an auxiliary output network to obtain an output of the auxiliary output network.

[0059] The electronic device inputs the auxiliary feature vector into the forward-looking gradient network 240. After obtaining the gradient feature vector, the gradient feature vector can be fused with the auxiliary feature vector to obtain a fused feature vector. As an example, the electronic device can add the gradient feature vector to each dimensional vector in the auxiliary feature vector to obtain a fused feature vector. In this way, the fused feature vector can also carry information about the contribution to the optimization target of the core network, so that the auxiliary network can biasedly optimize the optimization target of the core network.

[0060] S105. The electronic device updates the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

[0061] As an example, the optimization target of the auxiliary network may be the viewing time of the promoted content by the user, and the optimization target of the core network may be the conversion rate of the users converted by the promoted content.

[0062] It should be noted that the number of auxiliary networks is not limited to one, but may be multiple. Furthermore, the optimization target of each auxiliary network may be different. For example, the optimization target of other auxiliary networks may be users' forwarding, liking, and commenting on the promotional content.

[0063] In some examples, when the output of the auxiliary output network meets the optimization target of the auxiliary network or the optimization target of the core network, the sample feature vector is the feature vector corresponding to the positive sample; when the output of the auxiliary output network does not meet the optimization target of the auxiliary network and the optimization target of the core network, the sample feature vector is the feature vector corresponding to the negative sample. In other words, as long as the sample feature vector meets any one of the optimization targets of the auxiliary network and the core network, it is considered to be a positive sample. Not only can the number of positive and negative samples be balanced, but it can also further encourage the auxiliary network to optimize the core task (i.e., the optimization target of the core network) in a biased manner.

[0064] In some embodiments, after obtaining the output of the auxiliary output network, the electronic device can determine a first loss value based on the output of the auxiliary output network, the label value of the sample feature vector, and the objective function of the auxiliary network, and then update the weight of the auxiliary network based on the first loss value until the auxiliary network meets the first training stop condition.

[0065] It should be noted that the embodiment of the present disclosure does not specifically limit the first training stop condition, and those skilled in the art can set the first training stop condition according to actual needs. For example, the first training stop condition may be that the accuracy of the auxiliary network prediction is greater than a preset accuracy.

[0066] S106: The electronic device inputs the sample feature vector into the core exclusive network to obtain a core feature vector.

[0067] Similarly, the electronic device may input the sample feature vector into the shared network 210 to obtain the shared feature vector output by the shared network 210 , and then the electronic device may input the shared feature vector into the core exclusive network 250 to obtain the core feature vector.

[0068] S107. The electronic device inputs the core feature vector into the core output network to obtain the output of the core output network.

[0069] S108. The electronic device updates the weight of the core network according to the output of the core output network and the optimization target of the core network until the core network meets the second training stop condition.

[0070] In some embodiments, after obtaining the output of the core output network, the electronic device can determine a second loss value based on the output of the core output network, the label value of the sample feature vector, and the objective function of the core network, and then update the weight of the core network based on the second loss value until the core network meets the second training stop condition.

[0071] It should be noted that the embodiment of the present disclosure does not specifically limit the second training stop condition, and those skilled in the art can set the second training stop condition according to actual needs. For example, the second training stop condition may be that the accuracy of the core network prediction is greater than a preset accuracy.

[0072] The embodiments of the present disclosure do not limit the execution order of the above S102-S105 and S106-S108. In other embodiments, the electronic device may first execute S106-S108 and then execute S102-S105; or execute S102-S105 and S106-S108 in parallel.

[0073] Based on the above description, the present disclosure provides a multi-task model training method, which includes: first obtaining a sample feature vector, which is used to characterize the user's behavior on the promotion content; then inputting the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector, inputting the auxiliary feature vector into the forward-looking gradient network to obtain a gradient feature vector, which represents the contribution of the auxiliary feature vector to the optimization target of the core network; then fusing the gradient feature vector with the auxiliary feature vector and inputting it into the auxiliary output network. Since the input of the auxiliary output network includes the contribution to the optimization target of the core network, the auxiliary network can also optimize the optimization target of the core network, so that the auxiliary network can biasedly optimize the core target of the core task. Finally, the weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network. If the output of the auxiliary output network satisfies any one of the optimization targets of the auxiliary network and the core network, the sample feature vector is considered to be the feature vector corresponding to the positive sample, so that the auxiliary network can further biasedly optimize the optimization target of the core network to meet business needs.

[0074] The present disclosure also provides a promotion content processing method, see Figure 3 , which is a flow chart of a promotion content processing method provided by an embodiment of the present disclosure, the method comprising:

[0075] S301. The electronic device obtains a user feature vector for characterizing a user's behavior toward promotional content.

[0076] The user feature vector is similar to the sample feature vector, except that the user feature vector is the data used in this prediction, while the sample feature vector is the data used in the training phase. It should be noted that the electronic device needs to obtain the user's authorization in advance. After obtaining the user's authorization to use the corresponding data (such as the user's behavior on the promotion content mentioned above), the electronic device can obtain data such as the user's behavior on the promotion content.

[0077] S302: The electronic device obtains an inference result of the multi-task model according to the user feature vector and the multi-task model.

[0078] The multi-task model can be obtained through the multi-task model training method described above, which will not be described in detail here.

[0079] S303: The electronic device adjusts the promotion strategy for the promotion content according to the inference result.

[0080] In some examples, when the inference result indicates that the conversion rate of the promotional content is greater than the conversion rate threshold, and the playback time of the promotional content is greater than the time threshold, the electronic device increases the number of promotions of the promotional content; when the conversion rate of the promotional content is less than the conversion rate threshold, and the playback time of the promotional content is less than the time threshold, the electronic device reduces the number of promotions of the promotional content. In other examples, when the presentation type of the promotional content is a preset type (such as a video type), and the information of the promotional object in the promotional content is preset information (such as indicating that the promotional object is a game, a virtual item, or a physical object), the electronic device increases the number of promotions of the promotional content. In this way, the ineffective delivery of promotional content is reduced and the waste of resources is reduced.

[0081] The disclosed embodiment also provides a multi-task model training device, the multi-task model includes an auxiliary network, a core network and a look-ahead gradient network, the auxiliary network includes an auxiliary exclusive network and an auxiliary output network. Figure 4 As shown, the multi-task model training device 400 includes:

[0082] The sample acquisition module 401 is used to acquire a sample feature vector, wherein the sample feature vector is used to characterize the user's behavior towards the promotion content;

[0083] The vector processing module 402 is used to input the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector; input the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network; and fuse the gradient feature vector with the auxiliary feature vector and input the resultant into the auxiliary output network to obtain the output of the auxiliary output network;

[0084] The model training module 403 is used to update the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

[0085] Optionally, the look-ahead gradient network is updated based on an optimization objective of the core network.

[0086] Optionally, the core network includes a core exclusive network and a core output network; the vector processing module 402 is further used to input the sample feature vector into the core exclusive network to obtain a core feature vector; input the core feature vector into the core output network to obtain the output of the core output network; the model training module 403 is further used to update the weight of the core network according to the output of the core output network and the optimization target of the core network until the core network meets the second training stop condition.

[0087] Optionally, the vector processing module 402 is also used to input the sample feature vector into a shared network to obtain a shared feature vector; the vector processing module 402 is specifically used to input the shared feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector, and to input the shared feature vector into the core exclusive network to obtain a core feature vector.

[0088] Optionally, when the output of the auxiliary output network meets the optimization target of the auxiliary network or the optimization target of the core network, the sample feature vector is the feature vector corresponding to the positive sample; when the output of the auxiliary output network does not meet the optimization target of the auxiliary network and the optimization target of the core network, the sample feature vector is the feature vector corresponding to the negative sample.

[0089] The disclosed embodiment also provides a promotion content processing device, such as Figure 5 , the promotion content processing device 500 includes:

[0090] The acquisition module 501 is used to acquire a user feature vector for characterizing the user's behavior towards the promotion content.

[0091] The reasoning module 502 is used to obtain the reasoning result of the multi-task model according to the user feature vector and the multi-task model; the multi-task model is obtained by the multi-task model training method mentioned above.

[0092] The processing module 503 is used to adjust the promotion strategy for the promotion content according to the inference result.

[0093] The functions of the above modules have been described in detail in the method steps in the previous embodiment and will not be repeated here.

[0094] Reference below Figure 6 , which shows a schematic diagram of the structure of an electronic device 600 suitable for implementing the embodiment of the present disclosure, Figure 6 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0095] like Figure 6 As shown, the electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the electronic device 600 are also stored. The processing device 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0096] Typically, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device 600 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 6 The electronic device 600 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead.

[0097] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a non-transitory computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0098] It should be noted that the computer-readable medium disclosed above may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The 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 of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination of the above. The computer readable signal medium may also be any computer readable medium other than a computer readable storage medium, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0099] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0100] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device.

[0101] The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: obtains a sample feature vector, which is used to characterize the user's behavior on the promotion content; inputs the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector; inputs the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, which characterizes the contribution of the auxiliary feature vector to the optimization target of the core network; after fusing the gradient feature vector with the auxiliary feature vector, inputs the auxiliary output network to obtain the output of the auxiliary output network; updates the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition; or,

[0102] Acquire a user feature vector for characterizing the user's behavior on the promoted content; obtain an inference result of the multi-task model based on the user feature vector and the multi-task model; and adjust a promotion strategy for the promoted content based on the inference result.

[0103] Computer program code for performing the operations of the present disclosure may be written in one or more programming languages ​​or a combination thereof, including, but not limited to, object-oriented programming languages, such as Java, Smalltalk, C++, and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0104] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0105] The modules involved in the embodiments described in the present disclosure may be implemented by software or hardware. The name of a module does not limit the module itself in some cases. For example, the first acquisition module may also be described as a "module for acquiring at least two Internet Protocol addresses".

[0106] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0107] In the context of the present disclosure, 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, device, or equipment. 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, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0108] According to one or more embodiments of the present disclosure, Example 1 provides a multi-task model training method, wherein the multi-task model includes an auxiliary network, a core network, and a look-ahead gradient network, wherein the auxiliary network includes an auxiliary exclusive network and an auxiliary output network; the method includes:

[0109] Obtaining a sample feature vector, where the sample feature vector is used to characterize the user's behavior towards the promotion content;

[0110] Inputting the sample feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector;

[0111] Inputting the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network;

[0112] After fusing the gradient feature vector and the auxiliary feature vector, the resultant is input into the auxiliary output network to obtain the output of the auxiliary output network;

[0113] The weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

[0114] According to one or more embodiments of the present disclosure, Example 2 provides the method of Example 1, wherein the look-ahead gradient network is updated based on the optimization target of the core network.

[0115] According to one or more embodiments of the present disclosure, Example 3 provides the method of Example 1 or 2, wherein the core network includes a core exclusive network and a core output network; and the method further includes:

[0116] Inputting the sample feature vector into the core exclusive network to obtain a core feature vector;

[0117] Inputting the core feature vector into the core output network to obtain the output of the core output network;

[0118] According to the output of the core output network and the optimization target of the core network, the weight of the core network is updated until the core network meets the second training stop condition.

[0119] According to one or more embodiments of the present disclosure, Example 4 provides the method of Example 3, wherein the method further includes:

[0120] Inputting the sample feature vector into a shared network to obtain a shared feature vector;

[0121] The step of inputting the sample feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector comprises:

[0122] Inputting the shared feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector;

[0123] The step of inputting the sample feature vector into the core exclusive network to obtain a core feature vector includes:

[0124] The shared feature vector is input into the core exclusive network to obtain a core feature vector.

[0125] According to one or more embodiments of the present disclosure, Example 5 provides the method of Example 1, wherein when the output of the auxiliary output network meets the optimization target of the auxiliary network or the optimization target of the core network, the sample feature vector is the feature vector corresponding to the positive sample; when the output of the auxiliary output network does not meet the optimization target of the auxiliary network and the optimization target of the core network, the sample feature vector is the feature vector corresponding to the negative sample.

[0126] According to one or more embodiments of the present disclosure, Example 6 provides the method of Example 1, wherein the auxiliary exclusive network includes a deep neural network, a convolutional neural network, or a self-attention network; and the shared network includes a deep neural network, a convolutional neural network, or a self-attention network.

[0127] According to one or more embodiments of the present disclosure, Example 7 provides the method of Example 1, wherein the optimization target of the core network includes the conversion rate of the users converted by the promoted content; and the optimization target of the auxiliary network includes the viewing time of the users for the promoted content.

[0128] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of disclosure involved in the present disclosure is not limited to the technical solutions formed by a specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above disclosed concept. For example, the above features are replaced with the technical features with similar functions disclosed in the present disclosure (but not limited to) by each other.

[0129] In addition, although each operation is described in a specific order, this should not be understood as requiring these operations to be performed in the specific order shown or in a sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although some specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Some features described in the context of a separate embodiment can also be implemented in a single embodiment in combination. On the contrary, the various features described in the context of a single embodiment can also be implemented in multiple embodiments individually or in any suitable sub-combination mode.

[0130] Although the subject matter has been described in language specific to structural features and / or method logic actions, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. On the contrary, the specific features and actions described above are merely example forms of implementing the claims. Regarding the device in the above embodiment, the specific manner in which each module performs the operation has been described in detail in the embodiment related to the method, and will not be elaborated here.

Claims

1. A multi-task model training method, characterized in that: The multi-task model includes an auxiliary network, a core network and a look-ahead gradient network, the auxiliary network includes an auxiliary exclusive network and an auxiliary output network; the method includes: Obtaining a sample feature vector, where the sample feature vector is used to characterize the user's behavior towards the promotion content; Inputting the sample feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector; Inputting the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network; After fusing the gradient feature vector and the auxiliary feature vector, the resultant is input into the auxiliary output network to obtain the output of the auxiliary output network; The weight of the auxiliary network is updated according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

2. The method according to claim 1, characterized in that The look-ahead gradient network is updated based on the optimization objective of the core network.

3. The method according to claim 1 or 2, characterized in that: The core network includes a core exclusive network and a core output network; the method further includes: Inputting the sample feature vector into the core exclusive network to obtain a core feature vector; Inputting the core feature vector into the core output network to obtain the output of the core output network; According to the output of the core output network and the optimization target of the core network, the weight of the core network is updated until the core network meets the second training stop condition.

4. The method according to claim 3, characterized in that The method further comprises: Inputting the sample feature vector into a shared network to obtain a shared feature vector; The step of inputting the sample feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector comprises: Inputting the shared feature vector into an auxiliary exclusive network to obtain an auxiliary feature vector; The step of inputting the sample feature vector into the core exclusive network to obtain a core feature vector includes: The shared feature vector is input into the core exclusive network to obtain a core feature vector.

5. The method according to claim 1, characterized in that: When the output of the auxiliary output network meets the optimization target of the auxiliary network or the optimization target of the core network, the sample feature vector is a feature vector corresponding to a positive sample; When the output of the auxiliary output network does not meet the optimization target of the auxiliary network and the optimization target of the core network, the sample feature vector is a feature vector corresponding to a negative sample.

6. The method according to claim 4, characterized in that The auxiliary 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.

7. The method according to claim 1, characterized in that The optimization target of the core network includes the conversion rate of the users converted by the promotion content; the optimization target of the auxiliary network includes the viewing time of the users for the promotion content.

8. A promotion content processing method, characterized in that: include: Obtaining a user feature vector for characterizing the user's behavior towards the promoted content; According to the user feature vector and the multi-task model, an inference result of the multi-task model is obtained; the multi-task model is obtained by the method described in any one of claims 1 to 7; According to the inference result, the promotion strategy for the promotion content is adjusted.

9. A multi-task model training device, characterized in that: The multi-task model includes an auxiliary network, a core network and a look-ahead gradient network, wherein the auxiliary network includes an auxiliary exclusive network and an auxiliary output network; and the device includes: A sample acquisition module, used to acquire a sample feature vector, wherein the sample feature vector is used to characterize the user's behavior towards the promotion content; A vector processing module is used to input the sample feature vector into the auxiliary exclusive network to obtain an auxiliary feature vector; input the auxiliary feature vector into the look-ahead gradient network to obtain a gradient feature vector, wherein the gradient feature vector represents the contribution of the auxiliary feature vector to the optimization target of the core network; after fusing the gradient feature vector with the auxiliary feature vector, input the result into the auxiliary output network to obtain the output of the auxiliary output network; A model training module is used to update the weight of the auxiliary network according to the output of the auxiliary output network and the optimization target of the auxiliary network and the optimization target of the core network until the auxiliary network meets the first training stop condition.

10. A promotion content processing device, characterized in that: include: An acquisition module, used to acquire a user feature vector for characterizing the user's behavior towards the promoted content; An inference module, used for obtaining an inference result of the multi-task model according to the user feature vector and the multi-task model; the multi-task model is obtained by the method described in any one of claims 1 to 7; A processing module is used to adjust the promotion strategy for the promotion content according to the inference result.

11. An electronic device, characterized in that: include: a storage device having a computer program stored thereon; A processing device, configured to execute the computer program in the storage device to implement the method according to any one of claims 1 to 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processing device, the method according to any one of claims 1 to 8 is implemented.

13. A computer program product, characterized in that When the computer program product is run on a computer, the computer is caused to execute the method according to any one of claims 1 to 8.

Images