CN117273821A

CN117273821A - Issuing method, training method and related device of electronic equity certificates

Info

Publication number: CN117273821A
Application number: CN202311550669.3A
Authority: CN
Inventors: 杨涛
Original assignee: Ali Health Technology Hangzhou Co ltd
Current assignee: Ali Health Technology Hangzhou Co ltd
Priority date: 2023-11-20
Filing date: 2023-11-20
Publication date: 2023-12-22
Anticipated expiration: 2043-11-20
Also published as: CN117273821B

Abstract

The embodiment of the specification provides a method for issuing an electronic equity certificate, a training method and a related device. The method comprises the following steps: acquiring user behavior data; the user behavior data indicates that a user generates specified behaviors aiming at commodity objects; inputting the user behavior data into a user tower in a double-tower model, so that the double-tower model outputs a growth probability value of commodity benefits corresponding to the commodity objects after issuing electronic equity vouchers to users corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model scale of the double-tower model is smaller than the model scale of the specified conversion rate model; and under the condition that the increase probability value meets the specified condition, issuing an electronic equity credential to a user account corresponding to the user behavior data. The occupation of computer power resources can be reduced.

Description

Issuing method, training method and related device of electronic equity certificates

Technical Field

The embodiments in the present specification relate to the field of internet technologies, and in particular, to a method for issuing an electronic rights and interests credential, a training method, and related devices.

Background

With the development of internet technology, people have grown accustomed to purchasing goods on online shopping platforms. Likewise, many merchants offer online stores on online shopping platforms to sell goods.

With the rapid development of online shopping platforms, the number of online stores is increasing. So that competition among a plurality of online stores is also becoming more and more intense. In order to sell more goods, some merchants are willing to offer a portion of the benefit to consumers, and the sales of goods in online stores is increased by issuing coupons to consumers, thus obtaining more exposure and flow of goods in online stores.

There are typically a large number of users on online shopping platforms who have difficulty issuing consumer coupons to the truly demanding user accounts without the appropriate consumer coupon issuing policies. Or after the merchant invests a large amount of financial resources, the corresponding product sales conversion is not brought.

At this time, in order to screen out target user accounts requiring coupons among a large number of user accounts, a technician of the online shopping platform applies a machine learning model to the conversion rate prediction field. That is, the machine learning model is trained through a large number of samples, so that the machine learning model can output the purchase probability of the product represented by the user purchase product feature information corresponding to the user behavior data aiming at the input user behavior data and product feature information. Therefore, the machine learning model can be used for screening out a large number of user accounts, the user accounts with high probability of purchasing products exist, and coupons of corresponding products can be issued to the user accounts, so that the sales of the products can be improved.

With the development of society, the accuracy of a machine learning model for predicting conversion rate is better and larger, and the occupied computational resources are more and more. Because computer hardware resources are limited, existing machine learning models for predicting conversion rates place a significant burden on online shopping platforms.

Disclosure of Invention

Various embodiments of the present disclosure provide a method for issuing an electronic equity certificate, a training method, and related devices, which reduce the use of computing resources in the issuing process of the electronic equity certificate to a certain extent.

One embodiment of the present specification provides a method for issuing an electronic equity credential, the method including: acquiring user behavior data; the user behavior data indicates that a user generates specified behaviors aiming at commodity objects; inputting the user behavior data into a user tower in a double-tower model, so that the double-tower model outputs a growth probability value of commodity benefits corresponding to the commodity objects after issuing electronic equity vouchers to users corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model scale of the double-tower model is smaller than the model scale of the specified conversion rate model; and under the condition that the increase probability value meets the specified condition, issuing an electronic equity credential to a user account corresponding to the user behavior data.

One embodiment of the present specification provides a joint training method for training a twin tower model as described above, the method comprising: acquiring sample data; wherein the sample data has a sample tag; inputting the sample data into a specified conversion rate model and a double-tower model respectively, so that the specified conversion rate model outputs a conversion probability value for the sample data, and outputting a growth probability value corresponding to the sample data by the double-tower model; constructing a result loss function according to the conversion probability value, the increase probability value and the sample label; correcting the double-tower model according to the result loss function; and the result loss function acts on the double-tower model to realize the result distillation learning of the double-tower model on the specified conversion rate model.

One embodiment of the present specification provides an electronic equity certificate issuing apparatus, including: the acquisition module is used for acquiring user behavior data; the user behavior data indicates that a user generates specified behaviors aiming at commodity objects; the input module is used for inputting the user behavior data into a user tower in a double-tower model so that the double-tower model outputs a growth probability value of commodity income corresponding to the commodity object after issuing an electronic equity voucher to a user corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model scale of the double-tower model is smaller than the model scale of the specified conversion rate model; and the issuing module is used for issuing the electronic equity certificate to the user account corresponding to the user behavior data under the condition that the increase probability value meets the specified condition.

One embodiment of the present specification provides a joint training apparatus for training the aforementioned twin tower model, the joint training apparatus comprising: the sample acquisition module is used for acquiring sample data; wherein the sample data has a sample tag; a sample input module for inputting the sample data into a specified conversion rate model and a double-tower model, respectively, such that the specified conversion rate model outputs a conversion probability value for the sample data, and an expected growth probability value corresponding to the sample data is output by the double-tower model; a loss construction module for constructing a resulting loss function from the conversion probability value, the growth probability value, and the sample tag; the correction module is used for correcting the double-tower model according to the result loss function; and the result loss function acts on the double-tower model to realize the result distillation learning of the double-tower model on the specified conversion rate model.

The present description embodiment proposes a computer device comprising a memory storing a computer program and a processor implementing the method according to the above embodiment when the processor executes the computer program.

The present description provides a computer-readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method described in the above embodiments.

According to the embodiments provided by the specification, the target user account with increased commodity benefit can be brought to a merchant after the electronic benefit voucher is issued from the huge user account in the online shopping platform by adopting the small-scale double-tower model, and the computer computing power resource use of the machine learning model can be reduced.

Drawings

Fig. 1 is a schematic diagram of a dual tower model according to an embodiment of the present disclosure.

FIG. 2 is a schematic diagram of an architecture of an electronic rights and interests credential issuing system provided in one embodiment of the present description.

FIG. 3 is a flow chart of a method for issuing electronic equity vouchers provided in one embodiment of the present description.

Fig. 4 is a schematic diagram of a combined training architecture provided in one embodiment of the present description.

Fig. 5 is a flowchart of a joint training method according to an embodiment of the present disclosure.

Fig. 6 is a schematic diagram of a joint training architecture provided in one embodiment of the present disclosure.

Fig. 7 is a schematic block diagram of an electronic rights and interests credential issuing device according to an embodiment of the present disclosure.

Fig. 8 is a schematic block diagram of a joint training device according to an embodiment of the present disclosure.

Fig. 9 is a schematic diagram of a computer device according to an embodiment of the present disclosure.

Detailed Description

The following description of the embodiments of the present application will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are only some, but not all, of the embodiments of the present application. All other embodiments, which can be made by those skilled in the art based on the embodiments herein without making any inventive effort, are intended to be within the scope of the present application.

In the description of the embodiments of the present application, it should be understood that the terms "first," "second," and the like are used for descriptive purposes only and are not to be construed as indicating or implying a relative importance or an implicit indication of the number of features indicated. Thus, a feature defining "a first" or "a second" may explicitly or implicitly include one or more of the described features. In the description of the embodiments of the present application, the meaning of "a plurality" is two or more, unless explicitly defined otherwise.

In the related art, the online shopping platform can screen out a target account from a large number of user accounts by using an artificial intelligence technology, and can issue electronic equity certificates to the target account.

In the e-commerce field, the conversion rate may be a ratio of the number of users of the pointer to the specified function after the specified function is used and before the specified function is used. Specifically, for a new user promotion function, a first number of newly added users after the new user promotion function is adopted, a second number of newly added users before the new user promotion function is adopted, and a ratio between the first number and the second number is used as a conversion rate of the new user promotion function.

Specifically, a technician builds a large amount of sample data based on a machine learning algorithm, and trains to obtain a conversion rate model. The conversion rate model can be applied to different usage scenarios according to different usage purposes and training samples. Specifically, for example, for a coupon issuing scenario, a conversion rate model may be used to screen a target user group, so that after the coupon is issued, sales of the merchandise may be improved to some extent.

With the development of online shopping platforms, the number of commodities and users sold in online shopping platforms also reaches a very huge scale. In addition, in order to improve the accuracy of the conversion rate model, the model scale of the conversion rate model is also larger and larger. Therefore, the conversion rate model occupies more and more huge computer power resources in the running process.

In order to reduce the occupation of computer power resources and also maintain the normal functions of the online shopping platform, the skilled person thinks that a small-scale double-tower model can be used to realize the corresponding functions. The dual column model may include one customer column and one product column. The user tower is used for inputting user behavior data, and the product tower is used for inputting product characteristic information. And (5) performing dot product through the output of the user tower and the product tower to obtain a final output result.

However, it is difficult to achieve the accuracy of the original large-scale conversion rate model of the double-tower model in a short period of time by training the double-tower model purely by using the constructed training samples. In order to quickly improve the effect of the double-tower model, a technician thinks of a mode of carrying out distillation learning on the double-tower model by using a larger-scale conversion rate model, so that the double-tower model learns the large-scale conversion rate model. Thus, the capacity of the double-tower model can be improved faster.

With the training process of realizing conversion rate prediction based on the double-tower model. The technician finds that different merchants in the network sales platform, after issuing coupons to target user groups screened according to the conversion model. There may be a case where the sales of the merchant's goods increases, but the yield of the merchant is rather decreased. After analysis, the skilled person further finds that the main focus of the conversion model when outputting the results is the conversion itself, which is also aimed at improving the conversion during the training of the conversion model. Therefore, the target user group obtained by directly using the conversion rate model for screening can possibly improve sales of commodities to a certain extent, but due to the fact that the yield degree of coupons per se is different, the profit space of each commodity per se is different, so that the influence of issuing coupons on the benefits of the sold commodities can possibly be different for different merchants and different commodities.

Further, some users purchase the merchandise represented by the merchandise object after browsing the merchandise object, whether or not a coupon is received. Some users, however, are prompted to purchase the merchandise represented by the merchandise object after receipt of the coupon. Some users will not purchase the merchandise represented by the merchandise object after receiving the coupon. For merchants in the online shopping platform, after receiving coupons, users who purchase goods are promoted to be target user groups.

In order to enable the double-tower model to be more in line with the use needs of merchants in the online shopping platform. The technician makes further improvements to the dual-tower model so that the dual-tower model can output a growth probability value that will result in a growth of commodity revenue after issuing electronic equity vouchers for the user information. Therefore, coupons are issued to the target user groups obtained through screening according to the double-tower model, and the benefits of merchants can be improved to a certain extent. It will be appreciated that there is also some attribution to the increasing probability value output by the dual tower model. The electronic equity vouchers similar to coupons can be issued as an active intervention measure, and the degree of action of taking the active intervention measure on the commodity browsing to the final conversion into the commodity sales order can be represented by the increasing probability value output by the double-tower model for the commodity browsing to the final conversion into the commodity sales order. In the case of higher growth probability values, the conversion of browsing the commodity to the commodity sales order, which brings about the increase in commodity income, can be considered to be largely due to the active intervention taken. In the case that the increase probability value is smaller, it may be considered that after the electronic equity voucher is issued to the corresponding user account, it is difficult to cause increase in the commodity benefit, for example, issuing the electronic equity voucher does not cause the corresponding user account to indicate that the user purchases the corresponding commodity, or after issuing the electronic equity voucher, it may cause the corresponding user account to indicate that the user purchases the corresponding commodity, which may cause the merchant to have low or no benefit, or even negative benefit, due to the sale.

Because the purpose of use of the double tower model has changed, it is necessary to prepare a large number of training samples again and to exercise a large number of times. The double tower model can achieve good performance. Because of the small model size of the twin tower model, a great deal of training is required to obtain better performance. At this time, the technical staff find that the accuracy of the prediction of the gain growth probability of the double-tower model is also greatly improved by still adopting a larger-scale conversion rate model to distill and train the double-tower model through research. Namely, as knowledge in a conversion rate model for predicting conversion rate, the method has a good promotion effect on a double-tower model for predicting the growth probability of benefits. The improved double-tower model can be used for distillation learning to the conversion rate model, so that the training process of the double-tower model is accelerated, and the performance of the double-tower model is rapidly improved.

Referring to FIG. 1, one embodiment of the present description provides a dual tower model. The dual-tower model may include a user tower, a product tower, an interaction layer, and a dual-tower model output layer. The interaction layer is used for fusing the user characterization vector output by the user tower and the commodity characterization vector output by the product tower to form an interaction feature vector, and the double-tower model output layer outputs a result according to the interaction feature vector. The user tower may be used to input user behavior data and the product tower may be used to input merchandise characterization data. The consumer tower and the product tower include an input layer, an intermediate layer, and a tower output layer, respectively.

Referring to fig. 1 and 2, an example of an application scenario of an electronic rights and interests credential issuing system is provided in the present specification. The electronic equity voucher issuing system can provide an electronic equity voucher issuing method. The issuing system may include a client for receiving and presenting electronic equity vouchers, and a server deployed on an online shopping platform. The issuing system can screen out the issued electronic equity vouchers from the huge user accounts of the online shopping platform for the merchant when the merchant in the online shopping platform needs to issue the electronic equity vouchers, and can enable the commodity income of the merchant to be increased after the merchant screens out the target user accounts.

The dispensing system may be deployed with a dual tower model. The double tower model can be divided into a user tower and a product tower. The commodity characteristic data of the commodities in the online shopping platform can be input into the product tower in an offline state, so that commodity characterization vectors output by the product tower are obtained, and the commodity characterization vectors of each commodity are stored. Therefore, in the online use process, the stored commodity characterization vector can be directly used, and the occupation of computer power resources is reduced.

Further, the server of the online shopping platform may receive an item detail page access request sent by the user client, so as to request to display an item detail page of the corresponding item at the user client. The issuing system can acquire user behavior data of a user account corresponding to a client side sending an access request of an commodity detail page, acquire commodity representation vectors of commodities represented by the stored corresponding commodity detail page, input the user behavior data into a user tower to obtain user representation vectors of the corresponding user behavior data output by the user tower, and further input the user representation vectors and the commodity representation vectors into an interaction layer of the double-tower model and an output layer of the double-tower model to obtain a growth probability value.

The issuing system can identify whether the issuing of the electronic equity vouchers to the users corresponding to the user characterization vectors brings about the increase according to the increase probability values, and can issue the electronic equity vouchers to the users corresponding to the user characterization vectors when the identification brings about the increase of commodity benefits.

Because the whole process mainly uses the user tower with smaller scale to generate the user characterization vector for the user behavior data, and uses the output of the double-tower model output layer as the output result of the increase probability value, the whole issuing system of the electronic equity voucher has less occupied computer computing power resources in the online operation process.

In this scenario example, to do a promotional activity, a drug merchant C may issue electronic equity vouchers "New client 5-membered coupons" to some user accounts. Thus, the customer base of the drug merchant is augmented by issuing electronic equity vouchers "new customer 5-element coupons" to the user account. Further, a user a may be interested in purchasing some vitamins. The user A can operate the client to browse the commodity object list provided in the online shopping platform and browse the commodity detail page. The user A may browse the medicine X released by the medicine merchant C on the network sales platform, wherein the medicine X is a compound vitamin and the selling price is 135 yuan. User a may not have purchased a drug at drug merchant C such that user a belongs to a new customer relative to drug merchant C and the profit of drug merchant C for drug X may be 20 yuan. If user A purchases medicine X using the "New client 5-membered coupon", medicine merchant C will not have a loss due to the issuance of the electronic equity voucher.

However, the user a performs screening among the medicines provided by the plurality of medicine merchants, and does not necessarily exist in the explicit intention of the medicine merchant C to purchase the medicine X.

When the user A operates the client to browse the commodity detail page of the medicine X of the medicine merchant C, a commodity detail page access request for the medicine X is sent to the server first, so that the commodity detail page of the medicine X fed back by the server is obtained. At this time, the server may obtain user behavior data of the user a, and read a commodity characterization vector of the drug X from commodity characterization vectors generated by storing commodities corresponding to the product tower, input the user behavior data into a user tower in the double-tower model, generate a user characterization vector corresponding to the user behavior data by the user tower, input the user characterization vector and the commodity characterization vector into an interaction layer of the double-tower model and an output layer of the double-tower model, and obtain an expected growth probability value of 0.95. Assuming that the value of the specified threshold is 0.8, if the growth probability value is determined to be greater than the specified threshold, the electronic equity voucher may be considered to be issued. In this scenario example, the increase probability value output by the dual-tower model is 0.95 greater than the specified threshold, at which time the electronic equity voucher "new client 5-membered coupon" of the drug merchant C may be issued to the user a. That is, user A may purchase a drug at drug merchant C after receiving the electronic equity voucher "New client 5-membered coupon" in the client and use the electronic equity voucher "New client 5-membered coupon".

At this time, after the electronic equity certificate of "new client 5-element coupon" is issued, the medicine merchant C achieves the purpose of promoting new clients, and after the electronic equity certificate is issued, the increase of commodity income is brought.

Please refer to fig. 1. Embodiments of the present specification provide a system for issuing electronic equity vouchers. The issuing system of the electronic equity vouchers may include a client and a server. The client may be an electronic device with network access capabilities. Specifically, for example, the client may be a desktop computer, a tablet computer, a notebook computer, a smart phone, a digital assistant, a smart wearable device, a shopping guide terminal, a television, a smart speaker, a microphone, and the like. Wherein, intelligent wearable equipment includes but is not limited to intelligent bracelet, intelligent wrist-watch, intelligent glasses, intelligent helmet, intelligent necklace etc.. Alternatively, the client may be software capable of running in the electronic device. The server may be an electronic device with some arithmetic processing capability. Which may have a network communication module, a processor, memory, and the like. Of course, the server may also refer to software running in the electronic device. The server may also be a distributed server, and may be a system having a plurality of processors, memories, network communication modules, etc. operating in concert. Alternatively, the server may be a server cluster formed for several servers. Or, with the development of science and technology, the server may also be a new technical means capable of realizing the corresponding functions of the embodiment of the specification. For example, a new form of "server" based on quantum computing implementation may be possible.

Referring to fig. 3, an embodiment of the present disclosure provides a method for issuing electronic rights and interests credentials. The issuing method of the electronic rights and interests certificates can be applied to a server. The issuing method of the electronic equity certificate may include the following steps.

Step S101: acquiring user behavior data; the user behavior data indicates that the user generates specified behaviors aiming at commodity objects.

In some cases, a user may browse merchandise objects in the online shopping platform through a client to pick items of interest for purchase. In this process, the client may initiate an access request to the server according to a user operation, or perform operations such as collecting or joining in a shopping cart for some merchandise objects to represent the merchandise.

In this embodiment, the electronic equity certificate may be a certificate that exists in a digital drive for proving that it has a certain right. In the field of electronic commerce, electronic equity credentials may include, but are not limited to: discount coupons, no-freight coupons, new user coupons, full coupons, return coupons, oriented coupons, and the like.

In this embodiment, the commodity object is used to characterize the commodity in the form of a data object. It is understood that the merchandise object may be considered a data representation of the merchandise. The merchandise object may include a plurality of merchandise information. Specifically, the commodity information may be used to introduce a commodity characterized by the commodity object. For example, the form of merchandise information includes, but is not limited to, video, audio, images, text, or any combination. The merchandise information may include an introduction to a plurality of dimensions of the merchandise, such as function, structure, material, usage, appearance, and the like. The client can provide a commodity object list interface, and the commodity object list interface can comprise a plurality of showing pits, wherein each showing pit can show the thumbnail information of one commodity object. The thumbnail information may include partial commodity information of one commodity object. The client side can also display the commodity detail page of the commodity object under the operation of the user. The commodity information with more comprehensive commodity objects can be displayed in the commodity detail page.

In this embodiment, the user behavior data may be used to represent data generated due to operation behavior during the process of using the client by the user. Specifically, for example, the user behavior data may include: user search behavior data generated by keyword search and user click behavior data formed by clicking thumbnail information of commodity objects in a commodity object list interface are carried out; browsing behavior data for browsing the commodity detail page of the commodity object, collection behavior data for collecting the commodity object, shopping behavior data for adding the commodity object to the shopping cart, and the like.

In this embodiment, the user performs a specific action with respect to the commodity object, and it can be understood that an association relationship is established between the user account and the commodity object. The association relationship may indicate to some extent the degree to which the user is interested in the commodity object indicating the commodity.

In this embodiment, in the process of interaction between the client and the server, a corresponding request instruction is sent to the server according to the network communication protocol. The server may feed back the requested content to the client in response to the function of the request instruction. In this way, the server may analyze the user behavior data in accordance with the received request instructions. Of course, the collected and consolidated user behavior data provided by the third party software may also be received.

Step S102: inputting the user behavior data into a user tower in a double-tower model, so that the double-tower model outputs a growth probability value of commodity benefits corresponding to the commodity objects after issuing electronic equity vouchers to users corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model size of the dual column model is less than the model size of the specified conversion model.

In this embodiment, the dual-tower model may be trained in advance and used to output a growth probability value for the corresponding user and commodity object. The growth probability value may be used to represent a probability of an increase in commodity revenue due to a commodity sales caused by issuing an electronic equity voucher when the user purchases the commodity object to represent a commodity after issuing the electronic equity voucher to the user. Whether the user corresponding to the user behavior data can be suitable to be issued with the electronic equity credential can be determined by increasing the value of the probability value.

In this embodiment, the dual-tower model may be a smaller-scale machine learning model constructed based on deep learning. Specifically, the double tower model can be divided into a customer tower and a product tower. The user tower can output the user characterization vector after receiving the input user behavior data. The product tower may output the commodity characterization vector after receiving the input commodity characterization data. In some embodiments, the commodity characteristic data of the commodity object in the online shopping platform can be input into the product tower in an offline mode, and the commodity characterization vector output by the product tower corresponding to the commodity object is stored. In this way, in the subsequent use process, the user behavior data can be input into the user tower to obtain the user characterization vector, and after the commodity characterization vector of the commodity object is read, the final result can be output by the interaction layer and the double-tower model output layer. Of course, in some embodiments, the issuing method may also include a step of acquiring commodity feature data of the commodity object corresponding to the user behavior data, and generating, by the product tower, a commodity characterization vector corresponding to the commodity feature data. In the present embodiment, the commodity feature data of the commodity object may include at least part of commodity information of the commodity object. Such that the merchandise feature data may characterize the corresponding merchandise object.

Distillation learning is a machine learning method that is used primarily to compress and optimize large complex models for use in resource constrained environments. Distillation learning mainly uses a large complex model to guide a small lightweight model. Specifically, in this embodiment, the specified conversion rate model may be a large complex model, and the double-tower model may be a small lightweight model. In this way, knowledge learned by a given conversion model in a large number of sample exercises can be transferred to the double tower model by distillation learning.

In this embodiment, the two-tower model may be a training method for performing distillation learning on a specific conversion rate model in the training process. In this way, the dual-tower model is enabled to learn knowledge of a specified conversion model. In particular, the specified conversion rate model may be a machine learning model for predicting "conversion" after a specified behavior of a user with respect to a commodity object. Wherein, the "conversion" may include purchasing goods, registering account numbers, etc. In this embodiment, the specified conversion rate model may be a CVR (Conversion Rate) model. That is, the CVR model may be used to predict the probability of whether a user's click behavior against a merchandise object will "translate". As such, the specified conversion model may focus relatively on the occurrence of "conversions," such that knowledge about the "conversions" is learned from training sample data during training of the specified conversion model. As such, the specified conversion rate model may perform a prediction of whether "conversion" will occur for the input user behavior data and commodity feature data based on these learned knowledge. In the embodiment, the specified conversion rate model is learned through distillation of the double-tower model, so that the double-tower model can learn knowledge about conversion from the specified conversion rate model, and the double-tower model can predict the growth probability value on the basis of paying attention to conversion. Therefore, the double-tower model adopts a training mode of distillation learning to the specified conversion rate model, knowledge characteristics of the specified conversion rate model can be well utilized, and the capacity of the double-tower model is improved.

In this embodiment, the specified conversion rate model may be constructed based on a ranking algorithm in the machine learning field. Specifically, for example, machine learning algorithms that may implement a specified conversion model may include, but are not limited to, logistic regression, support vector machines, random forests, gradient-lifting trees, neural networks, and the like. In this embodiment, the specified conversion model may preferably be constructed using the DeepFM (Deep Factorization Machines) algorithm.

In this embodiment, the number of model parameters of the dual column model, and the computational resource requirements, may be less than the specified conversion model. In this way, in the process of executing the issuing method of the electronic equity certificate, the double-tower model is mainly adopted for operation, so that the occupation of computing resources is reduced.

In this embodiment, further, the probability of increase of the output of the dual tower model may also characterize the user who is taking the specified action, as to how sensitive to the intervention by the intervention. The degree of sensitivity to tampering may be used to indicate the likelihood that the corresponding user account will be translated into a purchase of the merchandise after the electronic equity voucher has been issued. The higher the level of sensitivity to tampering, the greater the likelihood that the user will be able to translate into a purchase of the merchandise after receiving the electronic equity voucher. The lower the level of sensitivity to tampering, the lower the likelihood of conversion to merchandise purchase after the user receives the electronic equity voucher may be considered. Further, it can be considered that the larger the increase probability value output by the double-tower model is, the larger the intervention sensitivity degree is, the smaller the increase probability value output by the double-tower model is, and the smaller the intervention sensitivity degree is.

Specifically, in some cases, the user may already have a category of goods for which it is clear that the user wants to purchase the goods, which is in a state of comparison or hesitation between the plurality of goods, or the plurality of merchants, that is, the user may be at a purchase intention critical point with respect to the goods of the plurality of merchants. The technician finds that the user at the critical point of purchase intent is more sensitive to the intervention by the intervention, i.e., if a merchant issues a coupon-like electronic equity voucher at that time, the user is prompted to purchase the merchandise using the coupon. Of course, the user may compare the commodities to multiple online shopping platforms and be at the purchase wish critical point, and at this time, a certain online shopping platform issues electronic equity vouchers similar to coupons to the user, or a merchant of a certain online shopping platform issues electronic equity vouchers similar to coupons to the user, which may prompt the user to decide to use the electronic equity vouchers to purchase the commodities. Thus, the user who is at the purchase intent critical point can be the target user who issues the electronic equity certificate. Therefore, the intervention sensitivity degree of the user who is issued the electronic equity certificate is represented by the increasing probability value output by the double-tower model, so that a target user account with higher intervention sensitivity degree can be screened out from huge user accounts by the double-tower model, and the electronic equity certificate can be accurately issued to the target user account. Of course, in some cases, the user may not have an explicit intent or tendency to purchase, but simply browse at will for some commodity objects of greater interest in the shopping network platform. For the part of users, users with higher intervention sensitivity can exist, so that the users with higher intervention sensitivity are screened out from the users without clear purchase intention through the increase probability value output by the double-tower model and can be used as target users issued with electronic equity vouchers, and the users with higher intervention sensitivity can be prompted to have purchase behaviors using the electronic equity vouchers. Of course, for some users, the electronic equity vouchers, even if received like coupons, are not translated into merchandise purchases. In the operation process based on the double-tower model, the corresponding increase probability value of the user insensitive to the intervention measures is also lower, and the electronic equity vouchers cannot be issued, so that the effective utilization rate of the electronic equity vouchers is improved to a certain extent.

Step S103: and under the condition that the increase probability value meets the specified condition, issuing an electronic equity credential to a user account corresponding to the user behavior data.

In this embodiment, whether to issue the electronic equity credential to the user account may be determined based on the value of the growth probability value. Specifically, a specified threshold may be set, and the electronic equity credential may be issued to the user account corresponding to the user behavior data when the growth probability value is greater than or equal to the specified threshold. In the event that the growth probability value is less than the specified threshold, no electronic equity credential is issued. Therefore, by using the double-tower model with smaller scale, the method can screen huge users in the online shopping platform, and obtain target users which are more likely to be converted into commodity income increase after issuing the electronic equity vouchers. The computational effort occupation of the online shopping platform is reduced, and meanwhile convenience is brought to merchants.

In some embodiments, a growth probability value corresponding to a user account with a specified behavior for each commodity object may be obtained corresponding to each commodity object, and the user accounts may be sorted according to the value of the growth probability value, and the first K user accounts may be selected to issue the electronic equity credential. K may be a specified positive integer.

Please refer to fig. 4 and 5. One embodiment of the present specification provides a joint training method that may be used to train the aforementioned twin tower model. The joint training method may include the following steps.

Step S201: acquiring sample data; wherein the sample data has a sample tag.

In this embodiment, sample data may be constructed based on history data. Specifically, two sets of sample data may be constructed based on the history data, and may be divided into a first set of sample data and a second set of sample data. Wherein the first set of sample data includes sample data that is issued with electronic rights credentials and the second set of sample data includes sample data that is not issued with electronic rights credentials. Further, the first set of sample data and the second set of sample data each include positive sample data and negative sample data, respectively. The sample tag has a first tag value that indicates that the sample data is a positive sample and a second tag value that indicates that the sample data is a negative sample. In this embodiment, by separating the first set of sample data that is issued with the electronic equity vouchers from the second set of sample data that is not issued with the electronic equity vouchers, the two-tower model can learn whether the electronic equity vouchers are issued or not during the training process, and can bring knowledge of the increase in commodity benefits to the end.

In this embodiment, the positive sample data in the first set of sample data includes user behavior data, merchandise feature data, and a first tag value. The positive sample data in the first set of sample data may indicate that the increase of commodity income brought by purchasing the commodity corresponding to the commodity feature data is increasing after the user corresponding to the user behavior data is issued the electronic equity voucher. Negative sample data in the first set of sample data includes user behavior data, merchandise feature data, and a second tag value. The negative sample data in the first set of sample data may indicate that the increase of commodity income brought by purchasing the commodity corresponding to the commodity feature data is negative increase after the user corresponding to the user behavior data is issued the electronic equity voucher.

Positive sample data in the second set of sample data includes user behavior data, merchandise feature data, and a first tag value. Positive sample data in the second set of sample data may represent a user corresponding to user behavior data, and purchase of a commodity corresponding to commodity feature data occurs. Negative sample data in the second set of sample data includes user behavior data, merchandise feature data, and a second tag value. The negative sample data in the second set of sample data may represent a user corresponding to the user behavior data not purchasing a commodity corresponding to the commodity feature data.

Step S203: the sample data is input into a specified conversion rate model and a double-tower model, respectively, such that the specified conversion rate model outputs a conversion probability value for the sample data, and a rise probability value corresponding to the sample data is output by the double-tower model.

In the present embodiment, the specified conversion rate model may be a machine learning model for which training has been completed in advance. In the combined training method, the specified conversion rate model is used as a teacher model, so that a double-tower model used as a student model learns knowledge of the specified conversion rate model through a distillation learning mode.

In the present embodiment, the user characteristic data and the commodity characteristic data in the sample data may be simultaneously input to the specified conversion rate model such that the specified conversion rate model outputs the conversion probability value. User characteristic data in the sample data may be input to the user towers in the dual tower model and commodity characteristic data may be input to the product towers in the dual tower model. And outputting the growth probability value by the double-tower model output layer of the double-tower model.

Step S205: and constructing a result loss function according to the conversion probability value, the increase probability value, the expected increase probability value and the sample label.

In this embodiment, the conversion probability value output by the specified conversion rate model may be used as a type of soft tag, and the increase probability value output by the double-tower model may be used as a soft prediction for predicting the soft tag. The knowledge of the specified conversion rate model may be learned by implementing the double-tower model by constructing the first residual information such that the increase probability value output by the double-tower model approximates the conversion probability value output by the specified conversion rate model. Specifically, a conversion probability value is generated by a specified conversion rate model under the temperature parameter T, and an increase probability value is generated by a double tower model under the same temperature parameter T.

Specifically, a resulting loss function l=αl can be constructed _soft +βL _hard . Wherein L is _soft Can be used to represent the first residual information, L _hard May be used to represent the second residual information. Alpha and beta may be used to represent weights of the first residual information and the second residual information, respectively.

In particular, the method comprises the steps of,。/>，/>。

in particular, the method comprises the steps of,. Wherein (1)>。

Where j may be used to represent a class j class classification tag,conversion profile representing output of a specified conversion modelThe value of the rate value is calculated,representing the increasing probability value of the output of the dual tower model, N represents the total number of class labels, T represents the temperature parameter, and K may be used to represent class K class labels. Wherein (1) >Value for representing sample tag, +.>For representing the probability of increase of the output of the double tower model, N for representing the number of samples and T for representing the temperature parameter.

Step S207: correcting the double-tower model according to the result loss function; and the result loss function acts on the double-tower model to realize the result distillation learning of the double-tower model on the specified conversion rate model.

In this embodiment, the two-tower model may be counter-propagated based on the loss value obtained by the result loss function operation, so as to update parameters included in the two-tower model, thereby implementing correction of the two-tower model. Further, the result loss function includes taking the output of the specified conversion rate model as a soft tag and taking the output of the double-tower model as first residual information formed by soft prediction, so that the result distillation learning from the double-tower model to the specified conversion rate model is also realized in the process of back propagation of the double-tower model.

In some embodiments, the specified conversion model includes a plurality of conversion model interlayers; the customer tower in the dual tower model includes a plurality of customer tower interlayers, and the product tower includes a plurality of product tower interlayers; the joint training method may further include: constructing at least one conversion model intermediate layer, and an intermediate layer loss function of the intermediate layer of the user tower and the intermediate layer of the product tower; and correcting a user tower middle layer and a product tower middle layer in the double tower model according to the middle layer loss function so as to realize middle layer distillation learning of the double tower model on the specified conversion rate model.

In the present embodiment, the data processing capability of the intermediate layer of the double-tower model can be enhanced by adopting a method of performing distillation learning from the intermediate layer of the double-tower model to the intermediate layer of the specified conversion rate model.

In this embodiment, the specified conversion model may have a plurality of conversion model interlayers. Accordingly, the customer tower and the product tower in the dual tower model may also each include a plurality of intermediate layers. Specifically, for example, a consumer tower may include a plurality of consumer tower interlayers and a product tower may include a plurality of product tower interlayers. The intermediate layer loss functions of the intermediate layers of the conversion rate model and the intermediate layers of the double-tower model can be respectively constructed so as to take the vector output by the intermediate layer of the conversion model in the designated conversion rate model as a learning target of the intermediate layer of the double-tower model, so that the vector output by the intermediate layer of the double-tower model approximates to the vector output by the intermediate layer of the conversion model.

In the present embodiment, the double-tower model is subjected to result distillation learning with respect to the specific conversion model, and further, the double-tower model is subjected to intermediate layer distillation learning with respect to the specific conversion model. The double-tower model can learn the knowledge of the specified conversion rate model more comprehensively, and the accuracy of the output result of the double-tower model is improved.

In some implementations, the conversion model middle layer includes a first vector generation layer for generating a first user feature vector corresponding to user behavior data of the sample data, and generating a first commodity feature vector corresponding to commodity feature data of the sample data; the user tower intermediate layer includes a second vector generation layer for generating a second user feature vector corresponding to user behavior data of the sample data, and the product tower intermediate layer includes a third vector generation layer for generating a second commodity feature vector corresponding to commodity feature data of the sample data; the dimension of the second user feature vector is smaller than the dimension of the first user feature vector, and the dimension of the second commodity feature vector is smaller than the dimension of the first commodity feature vector; the joint training method may further include: converting the first user feature vector of the specified conversion rate model into the same vector space as the second user feature vector to obtain a converted user feature vector, and converting the first commodity feature vector of the specified conversion rate model into the same vector space as the second commodity feature vector to obtain a converted commodity feature vector; constructing the middle layer loss function according to the converted user feature vector, the second user feature vector, the converted commodity feature vector and the second commodity feature vector; correcting the second and third vector producing layers of the dual column model according to the intermediate layer loss function.

In this embodiment, the conversion model middle layer, the user tower middle layer, and the product tower middle layer may all be hidden layers. Correspondingly, the first user feature vector, the first commodity feature vector, the second user feature vector and the second commodity feature vector can be hidden layer vectors. In this embodiment, the first user feature vector and the first commodity feature vector output by the intermediate layer of the conversion rate model have higher data dimensions. In order to enable the user tower middle layer and the product tower middle layer to learn from the conversion rate model middle layer, the first user feature vector and the first commodity feature vector output by the conversion rate model middle layer are both transferred to a vector space with the same second user feature vector and the second commodity feature vector, and the converted user feature vector and the converted commodity feature vector are obtained. The first user characteristic vector and the second commodity characteristic vector of the conversion rate model are converted into vector spaces corresponding to the double-tower model, so that the converted user characteristic vector and the converted commodity characteristic vector can better keep the knowledge of the conversion rate model, and the double-tower model can learn and understand the knowledge carried by the converted user characteristic vector and the converted commodity characteristic vector in the training process. Further, the converted user feature vector may be a learning target of the user tower middle layer, and the converted commodity feature vector may be a learning target of the product tower middle layer.

Specifically, the intermediate layer loss function may include:。

wherein,a second commodity feature vector for representing the output of a third vector-generating layer in the product column, +.>And the first commodity characteristic vector is used for representing the generation of the commodity characteristic data corresponding to the first vector generation layer in the specified conversion rate model.Second user feature vector for representing output of second vector generation layer in user tower in double tower model,/for the user tower>A first user feature vector representing the generation of corresponding user behavior data for a first vector generation layer in a specified conversion model.

In some embodiments. The specified conversion rate model comprises a conversion rate model output layer for outputting the conversion probability value; the conversion rate model middle layer comprises a final middle layer adjacent to the conversion rate model output layer, and the final middle layer outputs a final middle layer vector; the double-tower model is provided with a double-tower model output layer and an interaction layer adjacent to the double-tower model output layer, and the interaction layer outputs a fused interaction feature vector according to the user feature vector output by the user tower and the commodity feature vector output by the product tower; the method further comprises the steps of: projecting the interaction feature vector to a vector space of the final stage middle layer vector to obtain a projected interaction feature vector; the vector dimension of the interaction feature vector is smaller than the vector dimension of the final-stage middle layer vector; constructing an interaction layer loss function according to the projection interaction feature vector and the final intermediate layer vector; and correcting the interaction layer of the double-tower model according to the interaction layer loss function so as to realize interaction layer distillation learning of the double-tower model on the specified conversion rate model.

In this embodiment, after the result distillation learning and the middle layer distillation learning are performed on the double-tower model, the double-tower model may have good performance for the training sample. Further, in order to avoid overfitting of the double-tower model and improve generalization capability of the double-tower model, one-time distillation learning can be added to the double-tower model.

In this embodiment, the final intermediate layer of the specified conversion model may be the last hidden layer. The final intermediate layer vector output by the final intermediate layer is used for being input to the conversion rate model output layer. In this embodiment, the interaction layer of the dual-tower model may be used to fuse the user feature vector output by the user tower and the commodity feature vector output by the product tower, so as to obtain the fused interaction feature vector. The interaction feature vector may be for input to the dual-tower model output layer for the dual-tower model output layer to output results.

In this embodiment, the vector dimension of the interaction feature vector of the double-tower model is smaller than the vector dimension of the final intermediate layer vector of the specified conversion rate model. After the interaction feature vector of the double-tower model is projected to the vector space of the final-stage middle layer vector of the appointed conversion rate model, an interaction layer loss function is constructed according to the projected interaction feature vector and the final-stage middle layer vector, and the interaction layer of the double-tower model is corrected according to the interaction layer loss function, so that the generalization capability of the double-tower model can be enhanced, and the double-tower model is prevented from being over-fitted after the previous two times of distillation learning. Thus, the double-tower model after three distillation learning has relatively balanced performance.

In some embodiments, please refer to fig. 6. Projecting the interaction feature vector to a vector space of the final intermediate layer vector to obtain a projected interaction feature vector may include: respectively inputting the interaction feature vectors into a plurality of different projectors to obtain primary projection interaction feature vectors respectively output by the different projectors; wherein different projectors are respectively used for projecting the input characteristic vectors to the vector space of the final stage middle layer vector; and carrying out averaging processing on the primary projection interaction feature vectors output by different projectors to obtain the projection interaction feature vectors.

In this embodiment, a plurality of different projectors may be used to project interaction feature vectors to the vector space of the final intermediate layer vector, respectively. The plurality of projectors may be respectively constructed based on different projection algorithms such that primary projected interaction feature vectors generated by different projectors for the interaction feature vectors may be different. Therefore, the projection interaction feature vectors are obtained by carrying out the averaging processing on the plurality of primary projection interaction feature vectors, and overfitting can be prevented to a certain extent. Specifically, for example, the number of projectors may be 3. Of course, the number of the projectors may be 4 or 5, and is not limited.

Specifically, an interaction layer loss function may be constructed:。

wherein,q can be the number of projectors, b can be the sample size, t can be the final intermediate layer of the specified conversion model, s can be the interaction layer of the double-tower model,/-A>It may be a light source such as a projector,may be an average of a plurality of projectors and i may represent a batch.

Please refer to fig. 7. One embodiment of the present specification provides an electronic equity voucher issuing apparatus. The dispensing device includes: the system comprises an acquisition module, an input module and a release module.

The acquisition module is used for acquiring user behavior data; the user behavior data indicates that the user generates specified behaviors aiming at commodity objects.

The input module is used for inputting the user behavior data into a user tower in a double-tower model so that the double-tower model outputs a growth probability value of commodity benefits corresponding to the commodity object after issuing electronic rights and interests credentials to a user corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model size of the dual column model is less than the model size of the specified conversion model.

And the issuing module is used for issuing the electronic equity certificate to the user account corresponding to the user behavior data under the condition that the increase probability value meets the specified condition.

Please refer to fig. 8. One embodiment of the present specification provides a joint training apparatus for training the above-described twin tower model. The joint training device comprises: the device comprises a sample acquisition module, a sample input module, a residual error construction module and a correction module.

The sample acquisition module is used for acquiring sample data; wherein the sample data has a sample tag.

A sample input module for inputting the sample data into a specified conversion rate model and a double-tower model, respectively, such that the specified conversion rate model outputs a conversion probability value for the sample data, and an expected growth probability value corresponding to the sample data is output by the double-tower model.

And the loss construction module is used for constructing a result loss function according to the conversion probability value, the increase probability value and the sample label.

The correction module is used for correcting the double-tower model according to the result loss function; and the result loss function acts on the double-tower model to realize the result distillation learning of the double-tower model on the specified conversion rate model.

The present description also provides a computer-readable storage medium having stored thereon a computer program which, when executed by a computer, causes the computer to perform the method of any of the above embodiments.

The present description also provides a computer program product comprising instructions which, when executed by a computer, cause the computer to perform the method of any of the above embodiments.

Referring to fig. 9, the present description may provide a computer apparatus, including: a memory, and one or more processors communicatively coupled to the memory; the memory stores instructions executable by the one or more processors to cause the one or more processors to implement the method for processing an application service in any of the above embodiments.

In some embodiments, the electronic device may include a processor, a non-volatile storage medium, an internal memory, a communication interface, a display device, and an input device connected by a system bus. The non-volatile storage medium may store an operating system and associated computer programs.

User information or user account information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, etc.) referred to in various embodiments of the present description are information and data that are authorized by the user or sufficiently authorized by the parties, and the collection, use, and processing of relevant data requires compliance with relevant legal regulations and standards of the relevant countries and regions, and is provided with corresponding operation portals for the user to select authorization or denial.

It will be appreciated that the specific examples herein are intended only to assist those skilled in the art in better understanding the embodiments of the present disclosure and are not intended to limit the scope of the present invention.

It should be understood that, in various embodiments of the present disclosure, the sequence number of each process does not mean that the execution sequence of each process should be determined by the function and the internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

It will be appreciated that the various embodiments described in this specification may be implemented either alone or in combination, and are not limited in this regard.

Unless defined otherwise, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this specification belongs. The terminology used in the description is for the purpose of describing particular embodiments only and is not intended to limit the scope of the description. The term "and/or" as used in this specification includes any and all combinations of one or more of the associated listed items. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise.

It will be appreciated that the processor of the embodiments of the present description may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method embodiments may be implemented by integrated logic circuits of hardware in a processor or instructions in software form. The processor may be a general purpose processor, a Digital signal processor (Digital SignalProcessor, DSP), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), an off-the-shelf programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components. The methods, steps and logic blocks disclosed in the embodiments of the present specification may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of a method disclosed in connection with the embodiments of the present specification may be embodied directly in hardware, in a decoded processor, or in a combination of hardware and software modules in a decoded processor. The software modules may be located in a random access memory, flash memory, read only memory, programmable read only memory, or electrically erasable programmable memory, registers, etc. as well known in the art. The storage medium is located in a memory, and the processor reads the information in the memory and, in combination with its hardware, performs the steps of the above method.

It will be appreciated that the memory in the embodiments of this specification may be either volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile memory may be a read-only memory (ROM), a Programmable ROM (PROM), an Erasable Programmable ROM (EPROM), an Electrically Erasable Programmable ROM (EEPROM), or a flash memory, among others. The volatile memory may be Random Access Memory (RAM). It should be noted that the memory of the systems and methods described herein is intended to comprise, without being limited to, these and any other suitable types of memory.

Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein can be implemented as electronic hardware, or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present specification.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described system, apparatus and unit may refer to corresponding procedures in the foregoing method embodiments, and are not repeated herein.

In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods may be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of the units is merely a logical function division, and there may be additional divisions when actually implemented, e.g., multiple units or components may be combined or integrated into another system, or some features may be omitted or not performed. Alternatively, the coupling or direct coupling or communication connection shown or discussed with each other may be an indirect coupling or communication connection via some interfaces, devices or units, which may be in electrical, mechanical or other form.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiment.

In addition, each functional unit in each embodiment of the present specification may be integrated into one processing unit, each unit may exist alone physically, or two or more units may be integrated into one unit.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the present specification may be essentially or portions contributing to the prior art or portions of the technical solutions may be embodied in the form of a software product stored in a storage medium, including several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the embodiments of the present specification. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a read-only memory (ROM), a random-access memory (RAM), a magnetic disk, or an optical disk, etc.

The foregoing is merely specific embodiments of the present disclosure, but the scope of the present disclosure is not limited thereto, and any person skilled in the art can easily think about changes or substitutions within the technical scope disclosed in the present disclosure, and should be covered by the scope of the present disclosure. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method of issuing electronic equity vouchers, the method comprising:

acquiring user behavior data; the user behavior data indicates that a user generates specified behaviors aiming at commodity objects;

inputting the user behavior data into a user tower in a double-tower model, so that the double-tower model outputs a growth probability value of commodity benefits corresponding to the commodity objects after issuing electronic equity vouchers to users corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model scale of the double-tower model is smaller than the model scale of the specified conversion rate model;

and under the condition that the increase probability value meets the specified condition, issuing an electronic equity credential to a user account corresponding to the user behavior data.

2. A joint training method for training the twin tower model as in claim 1, the method comprising:

acquiring sample data; wherein the sample data has a sample tag;

inputting the sample data into a specified conversion rate model and a double-tower model respectively, so that the specified conversion rate model outputs a conversion probability value for the sample data, and outputting a growth probability value corresponding to the sample data by the double-tower model;

Constructing a result loss function according to the conversion probability value, the increase probability value and the sample label;

correcting the double-tower model according to the result loss function; and the result loss function acts on the double-tower model to realize the result distillation learning of the double-tower model on the specified conversion rate model.

3. The method of claim 2, wherein the sample data comprises a first set of sample data and a second set of sample data; the first set of sample data and the second set of sample data each include positive sample data and negative sample data, respectively; the sample data in the first group of sample data is issued electronic rights and vouchers, and the sample data in the second group of sample data is not issued electronic rights and vouchers.

4. The method of claim 2, wherein the specified conversion model comprises a plurality of conversion model interlayers; the customer tower in the dual tower model includes a plurality of customer tower interlayers, and the product tower includes a plurality of product tower interlayers; the method further comprises the steps of:

constructing at least one conversion model intermediate layer, and an intermediate layer loss function of the intermediate layer of the user tower and the intermediate layer of the product tower;

And correcting a user tower middle layer and a product tower middle layer in the double tower model according to the middle layer loss function so as to realize middle layer distillation learning of the double tower model on the specified conversion rate model.

5. The method of claim 4, wherein the conversion model middle layer comprises a first vector generation layer for generating a first user feature vector corresponding to user behavior data of the sample data, and generating a first commodity feature vector corresponding to commodity feature data of the sample data; the user tower intermediate layer includes a second vector generation layer for generating a second user feature vector corresponding to user behavior data of the sample data, and the product tower intermediate layer includes a third vector generation layer for generating a second commodity feature vector corresponding to commodity feature data of the sample data; the dimension of the second user feature vector is smaller than the dimension of the first user feature vector, and the dimension of the second commodity feature vector is smaller than the dimension of the first commodity feature vector; the method further comprises the steps of:

Converting the first user feature vector of the specified conversion rate model into the same vector space as the second user feature vector to obtain a converted user feature vector, and converting the first commodity feature vector of the specified conversion rate model into the same vector space as the second commodity feature vector to obtain a converted commodity feature vector;

constructing the middle layer loss function according to the converted user feature vector, the second user feature vector, the converted commodity feature vector and the second commodity feature vector;

correcting the second and third vector generation layers of the dual-tower model according to the intermediate layer loss function.

6. The method of claim 4, wherein the specified conversion rate model includes a conversion rate model output layer outputting the conversion probability value; the conversion rate model middle layer comprises a final middle layer adjacent to the conversion rate model output layer, and the final middle layer outputs a final middle layer vector; the double-tower model is provided with a double-tower model output layer and an interaction layer adjacent to the double-tower model output layer, and the interaction layer outputs a fused interaction feature vector according to the user feature vector output by the user tower and the commodity feature vector output by the product tower; the method further comprises the steps of:

Projecting the interaction feature vector to a vector space of the final stage middle layer vector to obtain a projected interaction feature vector; the vector dimension of the interaction feature vector is smaller than the vector dimension of the final-stage middle layer vector;

constructing an interaction layer loss function according to the projection interaction feature vector and the final intermediate layer vector;

and correcting the interaction layer of the double-tower model according to the interaction layer loss function so as to realize interaction layer distillation learning of the double-tower model on the specified conversion rate model.

7. The method of claim 6, wherein the step of projecting the interaction feature vector into the vector space of the final intermediate layer vector to obtain a projected interaction feature vector comprises:

respectively inputting the interaction feature vectors into a plurality of different projectors to obtain primary projection interaction feature vectors respectively output by the different projectors; wherein different projectors are respectively used for projecting the input characteristic vectors to the vector space of the final stage middle layer vector;

and carrying out averaging processing on the primary projection interaction feature vectors output by different projectors to obtain the projection interaction feature vectors.

8. An electronic rights voucher issuing apparatus, said issuing apparatus comprising:

the acquisition module is used for acquiring user behavior data; the user behavior data indicates that a user generates specified behaviors aiming at commodity objects;

the input module is used for inputting the user behavior data into a user tower in a double-tower model so that the double-tower model outputs a growth probability value of commodity income corresponding to the commodity object after issuing an electronic equity voucher to a user corresponding to the user behavior data; the double-tower model is obtained by distillation learning training with a specified conversion rate model; the model scale of the double-tower model is smaller than the model scale of the specified conversion rate model;

9. A co-training apparatus for training the twin tower model as in claim 1, the co-training apparatus comprising:

the sample acquisition module is used for acquiring sample data; wherein the sample data has a sample tag;

A sample input module for inputting the sample data into a specified conversion rate model and a double-tower model, respectively, such that the specified conversion rate model outputs a conversion probability value for the sample data, and an expected growth probability value corresponding to the sample data is output by the double-tower model;

a loss construction module for constructing a resulting loss function from the conversion probability value, the growth probability value, and the sample tag;

10. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the method of any one of claims 1 to 7 when executing the computer program.

11. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the method of any of claims 1 to 7.