Training method and device of business recommendation model, electronic equipment and storage medium

Abstract

The invention provides a business recommendation model training method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. In the application, after a training sample set comprising a first sample set for recommending a target service to a user and a second sample set for not recommending the target service to the user is obtained, gradient information obtained by training a service recommendation model based on the training sample set can be used for updating network parameters respectively corresponding to a plurality of networks included in the service recommendation model. Therefore, by introducing the super network capable of generating the network parameters corresponding to the first network and the second network respectively, the first sample set and the second sample set can be used for updating the network parameters corresponding to the first network and the second network respectively, so that the problem that data waste is caused by sharing information only in a shallow network stage is solved; the data size difference between the experimental group data and the control group data is large, and the experimental group data has selection deviation, so that the accuracy of searching the target user is improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a training method, device, electronic device and storage medium for a business recommendation model. Background Art

[0002] In various scenarios of daily life, before making service recommendations for users, it is usually necessary to mine potential users (i.e., target users) from multiple users in order to reduce operating costs.

[0003] For example, in a credit scenario, the more discount coupons are more attractive to users, and the more they can improve the credit rate (i.e., the probability of users taking out new loans). However, due to the high cost of coupons, it is impossible to issue large coupons to all users without limit. Therefore, it is necessary to find users (i.e., target users) who can maximize the credit rate to issue coupons.

[0004] Machine learning algorithms based on traditional associations can only simply associate user features with credit usage rates, but cannot estimate the extent to which credit usage will increase when issuing coupons compared to not issuing coupons, and thus cannot maximize the user's credit usage rate. In many cases, users who issue coupons themselves have a high credit usage rate. Causal inference algorithms can help to better find target users.

[0005] The current causal inference algorithms mainly include meta-learning-based causal inference algorithms and representation learning-based causal inference algorithms. Meta-learning-based causal inference algorithms include dual-model methods and single-model methods. However, for the dual-model method, both models require a large amount of experimental group (e.g., coupon-issuing users) data and control group (e.g., non-coupon-issuing users) data, and the two models cannot share information, resulting in data waste, while the single model cannot identify lower causal effects, such as the gain in user trust rate from coupon issuance. In addition, the experimental group data used in causal modeling is observational data rather than random data, and there is a large difference in the amount of data between the experimental group data and the control group data, which will affect the accuracy of the model.

[0006] Although the causal inference algorithm based on representation learning can share the information of the experimental group and the control group through a shared layer neural network, sharing information only in the shallow network stage (i.e., the shared layer neural network) will still cause data waste; and it still cannot solve the problem that the data volume gap between the experimental group data and the control group data is large and there is a bias in the selection of the experimental group data.

[0007] In view of this, how to improve the problem that only sharing information in the shallow network stage will still cause data waste, the large gap in data volume between the experimental group data and the control group data, and the selection bias in the experimental group data are issues that need to be urgently addressed. Summary of the invention

[0008] The embodiments of the present application provide a training method, device, electronic device and storage medium for a business recommendation model, which are used to improve the problem in the related technology that only sharing information in the shallow network stage still causes data waste, and the problem that the data volume difference between the experimental group data and the control group data is large and there is selection bias in the experimental group data, thereby improving the accuracy of finding target users.

[0009] In a first aspect, an embodiment of the present application provides a method for training a business recommendation model, the method comprising:

[0010] Acquire a training sample set of a service recommendation model; the training sample set includes: a first sample set for recommending a target service to a user and a second sample set for not recommending the target service to the user;

[0011] Based on the gradient information obtained by training the service recommendation model with the training sample set, the network parameters corresponding to the multiple networks included in the service recommendation model are updated; wherein the multiple networks include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network for generating the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the service usage rate after recommending the target service to the user, and the second network is used to determine the service usage rate after not recommending the target service to the user.

[0012] In an optional embodiment, obtaining a training sample set of a service recommendation model includes:

[0013] Obtaining user information corresponding to a plurality of users respectively; each piece of user information includes a plurality of user features of the corresponding user;

[0014] Based on a preset feature processing rule, feature dimension reduction processing is performed on multiple user information to obtain multiple user information after feature dimension reduction processing;

[0015] The multiple user information after feature dimension reduction is used as the training sample set.

[0016] In an optional embodiment, based on the gradient information obtained by training the service recommendation model with the training sample set, updating the network parameters corresponding to the multiple networks included in the service recommendation model includes:

[0017] Based on the label loss values ​​corresponding to the multiple training samples included in the training sample set, respectively, determining the gradient information corresponding to the shared network and the super network included in the multiple networks; wherein the shared network is used to extract shared features from the first sample set and the second sample set;

[0018] The network parameters of the shared network are updated based on the gradient information corresponding to the shared network, and the network parameters of the super network are updated based on the gradient information corresponding to the super network.

[0019] In an optional embodiment, before the label loss values ​​corresponding to the multiple training samples included in the training sample set, the method further includes:

[0020] For multiple training samples, perform the following operations respectively:

[0021] Extracting shared features from a training sample through a shared network to obtain a shared feature of the training sample;

[0022] Inputting the shared characteristics into the first network and the second network respectively, obtaining a first service usage rate output by the first network and a second service usage rate output by the second network;

[0023] Determine a prediction label of a training sample based on the first service usage rate and the second service usage rate;

[0024] The label loss value is obtained based on the predicted label and the true label of each training sample.

[0025] In an optional embodiment, determining a prediction label of a training sample based on the first service usage rate and the second service usage rate includes:

[0026] Based on the first service usage rate and the second service usage rate, obtaining a corresponding service usage rate difference;

[0027] When the service usage rate difference is greater than a set usage rate difference threshold, a predicted label of a training sample is determined to be a target user.

[0028] In an optional embodiment, updating the network parameters of the super network based on the gradient information corresponding to the super network includes:

[0029] Based on the preset neuron discarding algorithm, multiple neurons in the hypernetwork that participate in this model training are determined;

[0030] Update the network parameters of multiple neurons based on the gradient information corresponding to the hypernetwork.

[0031] In an optional embodiment, the output layer of the hypernetwork includes multiple neuron groups, and each neuron group is used to output part of the multiple parameters output by the hypernetwork.

[0032] In a second aspect, an embodiment of the present application further provides a training device for a business recommendation model, the device comprising:

[0033] A sample acquisition module is used to acquire a training sample set of a service recommendation model; the training sample set includes: a first sample set for recommending a target service to a user and a second sample set for not recommending a target service to a user;

[0034] A model training module is used to update the network parameters corresponding to the multiple networks included in the business recommendation model based on the gradient information obtained by training the business recommendation model with a training sample set; wherein the multiple networks at least include: a first network set for a first sample set and a second network set for a second sample set, and a super network that generates the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the business usage rate after recommending the target business to the user, and the second network is used to determine the business usage rate after not recommending the target business to the user.

[0035] In an optional embodiment, when acquiring a training sample set of a service recommendation model, the sample acquisition module is specifically used to:

[0036] Obtaining user information corresponding to a plurality of users respectively; each piece of user information includes a plurality of user features of the corresponding user;

[0037] Based on a preset feature processing rule, feature dimension reduction processing is performed on multiple user information to obtain multiple user information after feature dimension reduction processing;

[0038] The multiple user information after feature dimension reduction is used as the training sample set.

[0039] In an optional embodiment, when updating network parameters corresponding to a plurality of networks included in the business recommendation model based on the gradient information obtained by training the business recommendation model with the training sample set, the model training module is specifically used to:

[0040] Based on the label loss values ​​corresponding to the multiple training samples included in the training sample set, respectively, determining the gradient information corresponding to the shared network and the super network included in the multiple networks; wherein the shared network is used to extract shared features from the first sample set and the second sample set;

[0041] The network parameters of the shared network are updated based on the gradient information corresponding to the shared network, and the network parameters of the super network are updated based on the gradient information corresponding to the super network.

[0042] In an optional embodiment, before the label loss values ​​corresponding to the plurality of training samples included in the training sample set are respectively calculated, the model training module is further used to:

[0043] For multiple training samples, perform the following operations respectively:

[0044] Extracting shared features from a training sample through a shared network to obtain a shared feature of the training sample;

[0045] Inputting the shared characteristics into the first network and the second network respectively, obtaining a first service usage rate output by the first network and a second service usage rate output by the second network;

[0046] Determine a prediction label of a training sample based on the first service usage rate and the second service usage rate;

[0047] The label loss value is obtained based on the predicted label and the true label of each training sample.

[0048] In an optional embodiment, when determining a prediction label of a training sample based on the first service usage rate and the second service usage rate, the model training module is specifically used to:

[0049] Based on the first service usage rate and the second service usage rate, obtaining a corresponding service usage rate difference;

[0050] When the service usage rate difference is greater than a set usage rate difference threshold, a predicted label of a training sample is determined to be a target user.

[0051] In an optional embodiment, when updating the network parameters of the super network based on the gradient information corresponding to the super network, the model training module is specifically used to:

[0052] Based on the preset neuron discarding algorithm, multiple neurons in the hypernetwork that participate in this model training are determined;

[0053] Update the network parameters of multiple neurons based on the gradient information corresponding to the hypernetwork.

[0054] In a third aspect, an embodiment of the present application further provides an electronic device, including:

[0055] Processor; and

[0056] Memory for storing programs,

[0057] The program includes instructions, which, when executed by a processor, cause the processor to execute the training method for the business recommendation model as described in the first aspect.

[0058] In a fourth aspect, an embodiment of the present application further provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a computer to execute the training method of the business recommendation model as described in the first aspect.

[0059] In a fifth aspect, the present application provides a computer program product, which, when called by a computer, enables the computer to execute the steps of the training method for the business recommendation model as described in the first aspect.

[0060] The beneficial effects of this application are as follows:

[0061] In the training method of the service recommendation model provided in the embodiment of the present application, after obtaining a training sample set including a first sample set for recommending a target service to a user and a second sample set for not recommending the target service to the user, the network parameters corresponding to the multiple networks included in the service recommendation model can be updated based on the gradient information obtained by training the service recommendation model with the training sample set; wherein the aforementioned multiple networks may include at least: a first network set for the aforementioned first sample set and a second network set for the aforementioned second sample set, and a super network for generating network parameters corresponding to the first network and the second network, respectively; the first network can be used to determine the service usage rate after recommending the target service to the user, and the second network can be used to determine the service usage rate after not recommending the target service to the user.

[0062] In this way, by introducing a super network that can generate network parameters corresponding to the first network and the second network respectively, both the first sample set and the second sample set can be used to update the network parameters corresponding to the first network and the second network respectively, thereby improving the problem that only sharing information in the shallow network stage will still cause data waste, as well as the problem that the data volume difference between the experimental group data and the control group data is large and there is a selection bias in the experimental group data. Therefore, the accuracy of finding target users is improved.

[0063] In addition, other features and advantages of the present application will be described in the subsequent description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures particularly pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described here are used to provide a further understanding of the present application, constitute a part of the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0065] Figure 1 A schematic diagram of a specific scenario of user classification provided in an embodiment of the present application;

[0066] Figure 2 A schematic diagram of an optional system architecture applicable to the embodiments of the present application;

[0067] Figure 3 A schematic diagram of an implementation flow of a training method for a business recommendation model provided in an embodiment of the present application;

[0068] Figure 4 A schematic diagram of the structure of a business recommendation model provided in an embodiment of the present application;

[0069] Figure 5 A schematic diagram of an implementation flow of a method for determining a tag loss value provided in an embodiment of the present application;

[0070] Figure 6 A logical schematic diagram of a neuron discarding algorithm provided in an embodiment of the present application;

[0071] Figure 7 A method based on the embodiment of the present application is provided Figure 3 Schematic diagram of the

[0072] Figure 8 A schematic diagram of the structure of a training device for a business recommendation model provided in an embodiment of the present application;

[0073] Fig. 9 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0074] The embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present application are shown in the accompanying drawings, it should be understood that the present application can be implemented in various forms and should not be construed as being limited to the embodiments described herein. Instead, these embodiments are provided to provide a more thorough and complete understanding of the present application. It should be understood that the drawings and embodiments of the present application are only for exemplary purposes and are not intended to limit the scope of protection of the present application.

[0075] It should be understood that the various steps described in the method implementation of the present application can be performed in different orders and / or performed in parallel. In addition, the method implementation may include additional steps and / or omit the steps shown. The scope of the present application is not limited in this respect.

[0076] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to". The term "based on" means "based at least in part on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments". Relevant definitions of other terms will be given in the description below. It should be noted that the concepts of "first", "second", etc. mentioned in this application are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0077] It should be noted that the modifications of "one" and "plurality" mentioned in the present application are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0078] The names of the messages or information exchanged between multiple devices in the embodiments of the present application are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0079] Some of the terms used in the embodiments of the present application are explained below to facilitate understanding by those skilled in the art.

[0080] 1. Credit: refers to the user or object generating new borrowing or loan behavior.

[0081] 2. Meta-learning: focuses on how to learn learning strategies, that is, to improve the generalization ability of the model on limited data. It emphasizes learning how to learn, that is, by learning how to adjust the learning algorithm to adapt to new tasks or environments, and by simulating the human learning process, allowing machines to quickly adapt to new tasks and improve performance by learning learning strategies.

[0082] 3. Representation learning: focuses on how to use the model parameters, in what form and in what way to represent the model's input observation samples, aiming to find a better way to express the original data for subsequent task processing, such as classification. Representation learning improves the efficiency and accuracy of machine learning models by finding the best representation of data, and involves how to effectively represent data so that machine learning algorithms can better process and understand the data.

[0083] In short, meta-learning focuses on learning learning methods, that is, how to optimize the learning process itself, while representation learning focuses on learning data representation, that is, how to effectively express data for the processing of machine learning algorithms. These two methods each have their own application scenarios and advantages in the field of machine learning and artificial intelligence, and together promote the development of artificial intelligence technology.

[0084] Based on the above-mentioned nouns and related terminology explanations, the design concept of the embodiments of the present application is briefly introduced below:

[0085] In various scenarios of daily life, before recommending services to users, it is usually necessary to mine potential users (i.e., target users) from multiple users to reduce operating costs. It should be understood that the aforementioned potential users are users whose usage rate of the service will be greatly increased after the service is recommended to the user.

[0086] For example, in a credit scenario, by issuing discount coupons that reduce interest and fees, the attractiveness of related products can be increased, and the number of new and old users of credit can be increased. On the basis of increasing the number of users of credit, the coupons will also increase the user's loan amount and the number of loans, thereby increasing the profits of related institutions (such as companies). Therefore, it is very important to find a reasonable method of issuing coupons.

[0087] Intuitively, the greater the discount, the more attractive the coupon is to users and the more it can improve the credit rate. However, the cost of the coupon needs to be borne by the institution, so it is impossible to issue large coupons to all users without limit, which will cause huge losses. Taking the discount coupons with interest reduction as an example, the institution needs to find those users (i.e. target users) who can maximize the credit rate to issue coupons. Figure 1 The marketing quadrants shown in the figure show that users' responses to coupons can be divided into the following four categories:

[0088] (1) Marketing-sensitive groups: If coupons are issued, they will use the letter; if no coupons are issued, they will not use the letter.

[0089] (2) Natural conversion group: regardless of whether coupons are issued or not, they will use credit.

[0090] (3) Indifferent people: They will not use credit regardless of whether coupons are issued or not.

[0091] (4) Counter-effect group: People who don’t need to believe in coupons.

[0092] In order to maximize the efficiency of coupon issuance, how to accurately find the first type of marketing-sensitive population is the key to marketing coupon issuance. However, marketing algorithms based on traditional association machine learning can only simply associate user characteristics with credit rate, but cannot estimate the credit increase of issuing coupons compared to not issuing coupons, that is, coupon issuance gain, and thus cannot maximize the user's credit rate. In many cases, the person who issues coupons is already a person with a high credit rate. Therefore, causal inference marketing algorithms that can model causal effects (such as coupon issuance gain) are more suitable for coupon issuance scenarios and help to better find marketing-sensitive populations.

[0093] At present, the marketing coupon issuance algorithms based on causal inference mainly include causal inference algorithms based on meta-learning and causal inference algorithms based on representation learning:

[0094] (1) Causal inference algorithms based on meta-learning can include: dual-model (T-Learner) methods and single-model (S-Learner) methods. Since causal inference algorithms based on meta-learning are simple and easy to use, they are the most common methods used in the marketing field. The dual-model method is to use models to model the population that has received coupons and the population that has not received coupons respectively. When estimating, the data enters the two models, and the prediction results of the two models are used to make a difference to obtain the estimated increment. The advantage of this method is that the principle is intuitive and simple, and it can be implemented quickly. The single-model method takes whether a coupon is issued as a feature and adds it to the feature set to train a single model. The advantage is that it reduces the accumulation of errors between the two models and has a better recognition effect on weaker causal effects.

[0095] (2) The causal inference algorithm based on representation learning can fully integrate and utilize the information of the control group and the experimental group to obtain a more accurate causal model.

[0096] Using the above-mentioned causal inference algorithm, the coupon issuance data used in causal modeling is observational data rather than random data. Such data is not random but biased. There is also a situation where the amount of data in the experimental group and the control group is very different. For the dual model, modeling two sub-models requires that both the experimental group and the control group have enough data, and the two models cannot share information, resulting in data waste. For a single model, it is impossible to identify lower causal effects. Although the causal inference algorithm based on representation learning can share the information of the experimental group and the control group through a shared layer neural network, there is only preliminary information sharing in the shallow network stage, which cannot solve the problem of large data gaps and selection bias.

[0097] In view of this, in order to solve or improve the above-mentioned problems, an embodiment of the present application proposes a training method for a business recommendation model, which specifically includes: obtaining a training sample set of the business recommendation model; the training sample set includes: a first sample set for recommending a target business to a user and a second sample set for not recommending the target business to the user; based on the gradient information obtained by training the business recommendation model with the training sample set, updating the network parameters corresponding to the multiple networks included in the business recommendation model; wherein the multiple networks include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network for generating the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the business usage rate after recommending the target business to the user, and the second network is used to determine the business usage rate after not recommending the target business to the user.

[0098] In this way, by introducing a super network that can generate network parameters corresponding to the first network and the second network respectively, both the first sample set and the second sample set can be used to update the network parameters corresponding to the first network and the second network respectively, thereby improving the problem that only sharing information in the shallow network stage will still cause data waste, as well as the problem that the data volume difference between the experimental group data and the control group data is large and there is a selection bias in the experimental group data. Therefore, the accuracy of finding target users is improved.

[0099] In particular, the preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application, and the embodiments of the present application and the features in the embodiments may be combined with each other if there is no conflict.

[0100] See also Figure 2As shown, it is a schematic diagram of a system architecture applicable to an embodiment of the present application, and the system architecture may include: a terminal device (201a, 201b) and a server 202. The terminal device (201a, 201b) and the server 202 may exchange information through a communication network, wherein the communication mode adopted by the communication network may include: a wireless communication mode and a wired communication mode. Exemplarily, the terminal device (201a, 201b) may access the network through cellular mobile communication technology and communicate with the server 202. Wherein, the cellular mobile communication technology, for example, includes the fifth generation mobile communication (5th generation mobile networks, 5G) technology or the next generation mobile communication technology. Optionally, the terminal device (201a, 201b) may access the network through a short-range wireless communication mode and communicate with the server 202. Wherein, the short-range wireless communication mode, for example, includes wireless fidelity (wireless fidelity, Wi-Fi) technology.

[0101] The embodiment of the present application does not impose any restriction on the number of communication devices involved in the above system architecture. For example, the above system architecture may include more terminal devices, or fewer terminal devices, or other network devices. Figure 2 As shown, only the terminal device (201a, 101b) and the server 202 are described as examples, and the above communication devices and their respective functions are briefly introduced below.

[0102] The terminal device (201a, 201b) is a device that can provide voice and / or data connectivity to users, and can be a device that supports wired and / or wireless connection.

[0103] Exemplarily, the terminal devices (201a, 201b) may include, but are not limited to: mobile phones, tablet computers, laptop computers, PDAs, mobile internet devices (MID), wearable devices, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminal devices in industrial control, wireless terminal devices in unmanned driving, wireless terminal devices in smart grids, wireless terminal devices in transportation safety, wireless terminal devices in smart cities, or wireless terminal devices in smart homes, etc.

[0104] In addition, a related client may be installed on the terminal device (201a, 201b), and the client may be software, for example, an application (APP), a browser, a short video software, etc., or a web page, a mini-program, etc.; it should be noted that the terminal device (201a, 201b) in the embodiment of the present application may enable the above-mentioned client related to the training of the business recommendation model to send a training sample set for training the business recommendation model to the server 202, so as to perform subsequent method steps such as training of the business recommendation model.

[0105] Optionally, the terminal device (201a, 201b) can enable the above-mentioned client related to the training of the business recommendation model to send the user information of the current user to the server 202, so that the server 202 can subsequently determine whether the current user is the target user through the trained business recommendation model, and then determine whether it is necessary to recommend the target business to the current user.

[0106] For example, the server 202 determines whether the current user is a user whose usage rate will be greatly improved after the coupon is issued through the trained business recommendation model and the user information of the current user, and then determines whether to issue the coupon to the current user.

[0107] Server 202 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0108] It should be noted that if Figure 2 As shown, in an embodiment of the present application, a service recommendation model is deployed on the server 202. After the server 202 completes the training of the service recommendation model according to the training sample set, it can determine whether the current user is the target user and / or whether to recommend a service to the current user based on the obtained service recommendation model.

[0109] It is worth mentioning that in an embodiment of the present application, the server 202 can be used to obtain a training sample set of a service recommendation model, thereby updating the network parameters corresponding to the multiple networks included in the service recommendation model based on the gradient information obtained by training the service recommendation model with the training sample set. Among them, the aforementioned training sample set may include: a first sample set for recommending a target service to a user and a second sample set for not recommending a target service to a user. The aforementioned multiple networks may include at least: a first network set for a first sample set and a second network set for a second sample set, and a super network for generating network parameters corresponding to the first network and the second network, respectively. The first network can be used to determine the service usage rate after recommending the target service to the user, and the second network can be used to determine the service usage rate after not recommending the target service to the user.

[0110] The following describes the training method of the business recommendation model provided by the exemplary embodiment of the present application in combination with the above-mentioned system architecture and with reference to the accompanying drawings. It should be noted that the above-mentioned system architecture is only shown to facilitate understanding of the spirit and principles of the present application, and the implementation of the present application is not limited in this respect.

[0111] See also Figure 3 As shown, it is a schematic diagram of the implementation process of a training method for a business recommendation model provided in an embodiment of the present application. The execution subject takes a server as an example. The specific implementation process of the method is as follows:

[0112] S301: Obtain a training sample set for a business recommendation model.

[0113] The above training sample set may include: a first sample set for recommending target services to users and a second sample set for not recommending target services to users. Take the example of issuing coupons in a credit scenario as the target service. Recommending the target service to users means issuing coupons to users, and not recommending the target service to users means not issuing coupons to users. Therefore, the first sample set may be a set of user groups that issue coupons, and the second sample set may be a set of user groups that do not issue coupons.

[0114] The first sample set can be called experimental group data or coupon group data, and the second sample set can be called control group data or non-coupon group data. That is, the flow data of each coupon issuance (i.e., the flow distribution of each coupon issuance) can include two parts: experimental group data (or coupon group data) and control group data (or non-coupon group data). Of course, the first sample set and the second sample set can also have other names, which are not limited in the embodiments of the present application.

[0115] In an optional implementation, when executing step S301, the server may obtain or collect user information corresponding to multiple users, thereby performing feature dimension reduction processing on the multiple user information based on a preset feature processing rule, obtaining multiple user information after feature dimension reduction processing, and then using the multiple user information after feature dimension reduction processing as a training sample set. Among them, each user information includes multiple user features of the corresponding user. Exemplarily, the aforementioned user features may include the amount of provident fund payment, salary income, credit status, user gender, user age and other features.

[0116] In this way, at least one user feature related to the target business recommendation can be screened out from the multiple user features of each user. In this way, user features that are irrelevant to or have low relevance to the target business recommendation are removed, which not only improves the training efficiency of the model, reduces training time and memory usage, but also increases the generalization of the business recommendation model.

[0117] Taking the issuance of interest discount coupons on an Internet financial platform (i.e., an Internet finance platform) in a credit scenario as an example, the user information corresponding to multiple users obtained or collected may be: the user's credit report and its derivative data, Internet finance related data, user usage data of this platform, and the statistical characteristics of these data, etc., which is not limited to the embodiments of the present application.

[0118] The above-mentioned preset feature processing rules may include feature null value processing, feature stability processing and feature importance processing. Then the server can eliminate or remove null value features, unstable features and unimportant features in the user information according to the above-mentioned preset feature processing rules. Among them, the purpose of feature null value processing is to remove features with too large missing rates. These features have no discrimination for the business recommendation model. Removing these features can increase the model generalization of the business recommendation model. Feature stability processing refers to removing data with large distribution changes over time. These features will cause the model prediction results of the business recommendation model to vary greatly over time. After removal, the business recommendation model will be more stable. Feature importance processing refers to using a tree model to obtain the importance of features, removing features with zero importance, that is, features that do not participate in model training, and reducing the dimensionality of high-dimensional features is conducive to improving the training efficiency of the model and reducing training time and memory usage.

[0119] After the above-mentioned sample screening and feature processing, the high-dimensional features of the credit and user portraits of the original data (i.e., user information) can be reduced in dimension, and the most useful representative features can be extracted through feature processing methods. These features can better characterize the user portrait, thereby speeding up the training efficiency of the business recommendation model and improving the generalization ability of the model.

[0120] The multiple user information after the feature dimensionality reduction processing is the input feature or input data of the business recommendation model. According to the two stages of model training and model testing, the multiple user information after the feature dimensionality reduction processing can be divided into two parts: a sample training set and a sample test set. Exemplarily, 80% of the multiple user information after the feature dimensionality reduction processing is used as a sample training set, and 20% of the multiple user information after the feature dimensionality reduction processing is used as a sample test set. In the embodiment of the present application, the division ratio of the aforementioned sample training set and the aforementioned sample test set is not specifically limited and can be set according to actual conditions.

[0121] S302: Based on the gradient information obtained by training the service recommendation model with the training sample set, update the network parameters corresponding to the multiple networks included in the service recommendation model.

[0122] Among them, the above-mentioned multiple networks may include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network that generates network parameters corresponding to the first network and the second network respectively. In other words, the aforementioned super network is a neural network that generates network parameters (such as weights) for the aforementioned first network and the aforementioned second network. In addition, the first network can be used to determine the business utilization rate after recommending the target business to the user, and the second network can be used to determine the business utilization rate after not recommending the target business to the user. Taking the issuance of interest discount coupons on the mutual gold platform as an example, if the target user is a coupon issuing user, the coupon issuance utilization rate (i.e., credit possibility) of the target user can be determined through the first network. If the target user is a non-coupon issuing user, the credit possibility of the target user can be determined through the second network.

[0123] It should be noted that the first network and the second network may also be referred to as two branch neural networks of the business recommendation model. Of course, they may also have other names, which are not limited in the embodiments of the present application.

[0124] See also Figure 4As shown in the figure, in the training process of the neural network (that is, the learning network) in the business recommendation model, X is the processed features of the input sample (that is, the user features included in the user information after feature dimensionality reduction) g(X). First, the shared features of the experimental group data (that is, the training samples in the first sample set) and the control group data (that is, the training samples in the second sample set) are extracted through the shared network (that is, the shared layer neural network). Then, f1((g(X))) and f0((g(X))) are obtained respectively through two branch neural networks. Finally, the label loss value Y is calculated based on the predicted labels and the true labels obtained by f1((g(X))) and f0((g(X))). In the gradient back-propagation parameter update during the training process, the gradients generated by all the data can be used to update the shared network, while the experimental group data is used to update the first network f1, and the control group data is used to update the second network f0, thereby realizing the preliminary sharing of the experimental group data and the control group data in the shared network. However, at this time, the association information between the two groups of data cannot be shared in the first network and the second network, that is, the first network and the second network both update the network parameters according to the corresponding data, which will lead to a large difference in the amount of data between the experimental group data and the control group data (that is, the amount of data in the experimental group is small), resulting in insufficient learning of the business recommendation model for the experimental group data.

[0125] The above-mentioned shared network can be used to perform shared feature extraction on the first sample set and / or the second sample set. In other words, the feature X1 after the shared network extracts the features of the training samples in the first sample set can be input into the first network to obtain the first service usage rate f1((g(X1))), or input into the second network to obtain the second service usage rate f1((g(X1))); similarly, the feature X1 after the shared network extracts the features of the training samples in the second sample set can be input into the first network to obtain the first service usage rate f1((g(X2))), or input into the second network to obtain the second service usage rate f0((g(X2))).

[0126] It should also be noted that the above-mentioned representation learning network includes a shared network, a first network and a second network. Of course, it can also include other neural networks, such as a third network, etc., and the embodiments of the present application are not limited to this.

[0127] In order to solve the problem of insufficient information for representing learning shared experimental groups and data groups, e.g. Figure 4 As shown, the super network can be organically combined with the representation learning network. Therefore, in a possible implementation, when executing step S302, the server can determine the gradient information corresponding to the shared network and the super network included in the multiple networks based on the label loss values ​​corresponding to the multiple training samples included in the training sample set, thereby updating the network parameters of the shared network based on the gradient information corresponding to the shared network, and updating the network parameters of the super network based on the gradient information corresponding to the super network.

[0128] Based on the above method, in the training process of multiple networks (i.e., super networks - representation learning networks) in the business recommendation model, the super network h outputs the network parameters w1 of the first network f1 and the network parameters w0 of the second network f0, and the input is the mapping vector e of the branch network sequence number. i For example, the mapping vector of the first network f1 is e1 and the mapping vector of the second network f0 is e0. In each training of the business recommendation model, the gradient information generated by the experimental group data and the control group data is used to update the network parameters (such as hyperparameters) corresponding to the super network and the shared network respectively. Therefore, the entire representation learning network can dynamically share the information of the experimental group and the control group data, unlike the traditional representation learning model where only the shared network can share the experimental group and the control group data information, and the two branch networks f1 and f0 can only learn the information of the experimental data in this group, resulting in data waste, thereby avoiding the problem of deviation in the training of the first network and the second network due to the difference in data volume.

[0129] It should be noted that in the embodiments of the present application, the branch network or shared network can be replaced by any mature deep learning network, such as a residual neural network or a densely connected network, as long as it belongs to a supervised learning method, so as to flexibly complete model training and deployment, without specific limitation.

[0130] In an optional implementation, the server iteratively trains the service recommendation model based on the label loss values ​​corresponding to the multiple training samples included in the training sample set. In each iterative training process for the service recommendation model, for any one of the multiple training samples, i.e., the first training sample, refer to Figure 5 As shown, the servers perform the following operations respectively:

[0131] S501: extracting shared features from a first training sample through a shared network to obtain shared features of the first training sample.

[0132] The first training sample mentioned above can be a training sample in the first sample set (ie, experimental group data) or a training sample in the second sample set (ie, control group data). Furthermore, the shared feature mentioned above can be a feature that can be input into the first network or a feature that can be input into the second network training.

[0133] For example, assuming that the shared network is represented by g, the processed features of the first training sample, i.e., the input training sample, are represented by X i The shared features of the first training sample can be expressed as g(X i ) represents. Wherein, i represents that the first training sample is the i-th training sample in the aforementioned training sample set.

[0134] S502: Input the shared features into the first network and the second network respectively, and obtain a first service usage rate output by the first network and a second service usage rate output by the second network.

[0135] Exemplarily, when executing step S502, the server shares the feature g(X i ) are input into the first network f1 and the second network f0 respectively, then the first service usage rate output by the first network (that is, the service usage rate after the target service is recommended to the user corresponding to the first training sample) can be expressed as f1((g(X i ))), and the second service usage rate output by the second network (i.e., the service usage rate after not recommending the target service to the user corresponding to the first training sample) can be expressed as f0((g(X i ))).

[0136] S503: Determine a prediction label of a first training sample based on the first service usage rate and the second service usage rate.

[0137] In an optional implementation, when executing step S503, after obtaining the first service usage rate and the second service usage rate, the server can obtain the corresponding service usage rate difference based on the first service usage rate and the second service usage rate, so that when the service usage rate difference is greater than a set usage rate difference threshold, the predicted label of the first training sample is determined to be the target user. That is, the user corresponding to the first training sample is the user whose service usage rate increases significantly after the target service is recommended.

[0138] Taking the distribution of interest discount coupons on the mutual finance platform as an example, the server obtains the first service usage rate f1((g(X i ))) and the second service utilization rate f0((g(X i )))After that, the service utilization rate difference Δf can be obtained i =f1((g(X i )))-f0((g(X i ))). If we assume that the service utilization rate difference Δf i =97.2%-34.8%=62.4%, and the usage difference threshold is set to 50%, then it can be determined that the service usage difference 62.4% is greater than the set usage difference threshold 50%, and at this time it can be determined that the predicted label of the first training sample is the target user. It should be understood that the service usage difference Δf i That is, the user's sensitivity to coupons, which can also be expressed as effect i =f1((g(X i )))-f0((g(X i ))), effect iThe larger the value, the more sensitive the user is to coupons. Therefore, issuing coupons to relatively sensitive users can achieve better business recommendation results.

[0139] For another example, if we assume that the service usage rate difference Δf i =65.7%-48.2%=17.5%, and the set usage difference threshold is still 50%, then it can be determined that the service usage difference of 17.5% is less than the set usage difference threshold of 50%, and at this time it can be determined that the predicted label of the first training sample is not the target user. In other words, the user corresponding to the first training sample is not sensitive to coupons, that is, whether the coupon is issued to the user has little effect on whether the user uses new credit or the increase is small.

[0140] S504: Obtain a label loss value based on the predicted label and the true label of a training sample.

[0141] Exemplarily, when executing step S504, it is assumed that the predicted label is represented as The true label is represented as y0, and the label loss value can be expressed as: Furthermore, when the user's label is "1", it indicates that the user is the target user, and when the user's label is "0", it indicates that the user is not the target user. If the predicted label corresponding to the first training sample is The true label corresponding to the first training sample Then we can know the label loss value

[0142] It should be noted that in the embodiments of the present application, other methods (for example, logarithmic operations, etc.) can also be used to obtain label loss values ​​based on predicted labels and true labels, and there is no specific limitation on this.

[0143] After the server executes the method steps recorded in the above steps S501 to S504 for the above multiple training samples respectively, the label loss values ​​corresponding to the multiple training samples included in the training sample set can be obtained, thereby determining the gradient information corresponding to the shared network and the super network respectively, and then updating the network parameters corresponding to the shared network and the super network respectively.

[0144] In an optional implementation, when the server updates the network parameters of the supernetwork based on the gradient information corresponding to the supernetwork, it can determine multiple neurons in the supernetwork that participate in the current model training based on a preset neuron discarding algorithm (e.g., a temporary withdrawal or discarding (dropout) algorithm), thereby updating the network parameters of multiple neurons based on the gradient information corresponding to the supernetwork. In this way, during forward propagation, the activation values ​​of some neurons in the supernetwork are stopped with a certain probability (e.g., probability p). Since it will not be too dependent on certain local features, this can make the model more generalizable; and since the number of neurons in a single training process is reduced, the training speed of the supernetwork representation learning structure (i.e., the business recommendation model) is accelerated. Therefore, the training speed is improved while improving the robustness of the model.

[0145] For example, see Figure 6 As shown in , assume that the hypernetwork has a multilayer perceptron with 1 hidden layer and 5 hidden layer units. When the dropout algorithm is applied to the hidden layer, the hidden layer units are set to 0 with a probability of p, then the hidden layer can be regarded as a network containing only a subset of the original neurons. Figure 6 As shown in the figure, the dropout algorithm deletes the hidden layer units h2 and h5 in the hidden layer. Therefore, the calculation of the output no longer depends on h2 and h5, and their respective gradients will disappear when performing back propagation. In this way, the calculation of the output layer will not be overly dependent on any hidden layer unit in h1~h5.

[0146] It should be noted that the dropout algorithm is usually not used in the subsequent testing process. For the trained business recommendation model and a new sample, no nodes will be discarded, so standardization is not required.

[0147] In order to reduce the number of parameters in the training process, the output layer of the hypernetwork can include multiple neuron groups, each of which is used to output some of the multiple parameters output by the hypernetwork. That is, the output layer of the hypernetwork can be split into multiple neuron groups of equal length, each of which outputs a segment of the final output, and all output segments are finally spliced ​​into network parameters to reduce the number of parameters. For example, assuming that the number of neurons in the final output layer of the hypernetwork is 100 and the output parameter is n, 100*n parameters need to be output. If it is split into two segments, each segment uses 50 neurons to output the final total parameters of 100 / 2*n / 2=50*n.

[0148] Based on the method steps described in steps S301 to S302 above, refer to Figure 7As shown, the application scenario is the issuance of coupons on the Internet finance platform. For example, the traffic distribution for each coupon issuance can include two parts: the coupon issuance group (ie, the experimental group) and the non-coupon issuance group (ie, the control group). First, the data of the online coupon issuance group and the non-coupon issuance group are collected, and then the high-dimensional features of the credit investigation and user portrait (ie, user information) of the original data are reduced through sample screening and feature processing. The most useful representative features are extracted through feature processing methods. These features can better characterize the user portrait. The screened user feature data is divided into a training set and a test set, and training is performed on the training set to obtain a coupon issuance model (ie, a business recommendation model). The model is tested offline on another test data. The Gini score commonly used in causal algorithms can be used to measure the offline effect of the model. Under the same data, the higher the score, the better the model effect, and it is more likely to achieve better results online.

[0149] like Figure 7 As shown in the figure, after data collection and feature processing, the training of the hypernetwork representation learning model is connected. The best model is obtained through offline training and testing, and then the model is put online on the platform. The online model input is the characteristics of each user. Then the model scores each user online and issues coupons to users with high scores exceeding a certain threshold. These users are the highly sensitive users of coupons (i.e., target users) selected by the model. The probability of increasing credit by issuing coupons to these highly sensitive users is the highest compared to not issuing coupons, thereby achieving the effect of maximizing the gain from issuing coupons and improving the efficiency of issuing coupons and the platform's revenue.

[0150] In addition, as the time span increases, the distribution of user data will change, and the effect of training the model with old data will decrease. At this time, the online model needs to be updated. Figure 4 The model can be updated through the process.

[0151] Based on the above method, the platform revenue can be maximized, for example, the credit rate or loan increase can be maximized at a limited cost. This method avoids the drawback that the traditional association machine learning model cannot reveal this causal effect (i.e., coupon gain), and can best find marketing-sensitive people, thereby obtaining the maximum credit rate increment. In addition, this method solves the problem that causal meta-learning methods and representation learning methods cannot handle unbalanced experimental and control group data and the problem of biased experimental data. The hypernetwork representation learning model of the present application can make full use of the data of the experimental and control groups, not only sharing data information in the shared network, but also sharing the information of the experimental and control groups in the branch networks (i.e., the first network and the second network), solving the problem of biased experimental data and unbalanced data volume between the experimental and control groups.

[0152] In other words, in order to alleviate the problem of too large a data gap between the experimental group and the control group in causal modeling, as well as the problem of selection bias in the experimental group data, the hypernetwork is integrated into the existing representation learning method to maximize the improvement of user credit rate by issuing coupons. In both offline testing and online applications, better improvement effects were obtained than existing causal inference methods; at the same time, the model prediction results of the business recommendation model are also more robust.

[0153] In summary, in the training method of the business recommendation model provided in the embodiment of the present application, after obtaining a training sample set including a first sample set for recommending a target business to a user and a second sample set for not recommending a target business to a user, the gradient information obtained by training the business recommendation model based on the training sample set can be used to update the network parameters corresponding to the multiple networks included in the business recommendation model; wherein the aforementioned multiple networks at least include: a first network set for the first sample set and a second network set for the second sample set, and a super network for generating network parameters corresponding to the first network and the second network, respectively. In this way, by introducing a super network capable of generating network parameters corresponding to the first network and the second network, respectively, the first sample set and the second sample set can both be used to update the network parameters corresponding to the first network and the second network, respectively, thereby improving the problem that only sharing information in the shallow network stage will still cause data waste, and the data volume difference between the experimental group data and the control group data is large and there is a selection bias in the experimental group data, thereby improving the accuracy of finding target users.

[0154] Further, based on the same technical concept, the embodiment of the present application provides a training device for a business recommendation model, and the training device for a business recommendation model is used to implement the above method flow of the embodiment of the present application. Figure 8 As shown, the training device 800 of the business recommendation model includes: a sample acquisition module 801 and a model training module 802, wherein:

[0155] The sample acquisition module 801 is used to acquire a training sample set of a service recommendation model; the training sample set includes: a first sample set for recommending a target service to a user and a second sample set for not recommending a target service to a user;

[0156] The model training module 802 is used to update the network parameters corresponding to the multiple networks included in the business recommendation model based on the gradient information obtained by training the business recommendation model with the training sample set; wherein the multiple networks include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network that generates the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the business usage rate after recommending the target business to the user, and the second network is used to determine the business usage rate after not recommending the target business to the user.

[0157] In an optional embodiment, when acquiring a training sample set of a service recommendation model, the sample acquisition module 801 is specifically used to:

[0158] Obtaining user information corresponding to a plurality of users respectively; each piece of user information includes a plurality of user features of the corresponding user;

[0159] Based on a preset feature processing rule, feature dimension reduction processing is performed on multiple user information to obtain multiple user information after feature dimension reduction processing;

[0160] The multiple user information after feature dimension reduction is used as the training sample set.

[0161] In an optional embodiment, when updating network parameters corresponding to a plurality of networks included in the business recommendation model based on the gradient information obtained by training the business recommendation model with the training sample set, the model training module 802 is specifically used to:

[0162] Based on the label loss values ​​corresponding to the multiple training samples included in the training sample set, respectively, determining the gradient information corresponding to the shared network and the super network included in the multiple networks; wherein the shared network is used to extract shared features from the first sample set and the second sample set;

[0163] The network parameters of the shared network are updated based on the gradient information corresponding to the shared network, and the network parameters of the super network are updated based on the gradient information corresponding to the super network.

[0164] In an optional embodiment, before the label loss values ​​corresponding to the multiple training samples included in the training sample set are respectively calculated, the model training module 802 is further used to:

[0165] For multiple training samples, perform the following operations respectively:

[0166] Extracting shared features from a training sample through a shared network to obtain a shared feature of the training sample;

[0167] Inputting the shared characteristics into the first network and the second network respectively, obtaining a first service usage rate output by the first network and a second service usage rate output by the second network;

[0168] Determine a prediction label of a training sample based on the first service usage rate and the second service usage rate;

[0169] The label loss value is obtained based on the predicted label and the true label of each training sample.

[0170] In an optional embodiment, when determining a prediction label of a training sample based on the first service usage rate and the second service usage rate, the model training module 802 is specifically used to:

[0171] Based on the first service usage rate and the second service usage rate, obtaining a corresponding service usage rate difference;

[0172] When the service usage rate difference is greater than a set usage rate difference threshold, a predicted label of a training sample is determined to be a target user.

[0173] In an optional embodiment, when updating the network parameters of the super network based on the gradient information corresponding to the super network, the model training module 802 is specifically used to:

[0174] Based on the preset neuron discarding algorithm, multiple neurons in the hypernetwork that participate in this model training are determined;

[0175] Update the network parameters of multiple neurons based on the gradient information corresponding to the hypernetwork.

[0176] Based on the description of the above method embodiment and device embodiment, the exemplary embodiment of the present invention further provides an electronic device, including: at least one processor; and a memory connected to the at least one processor in communication. The memory stores a computer program that can be executed by the at least one processor, and the computer program is used to enable the electronic device to perform the method according to the embodiment of the present invention when executed by the at least one processor.

[0177] An embodiment of the present application also provides a non-transitory computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to execute a method according to an embodiment of the present application.

[0178] An embodiment of the present application also provides a computer program product, including a computer program, wherein the computer program, when executed by a processor of a computer, is used to cause the computer to execute a method according to an embodiment of the present application.

[0179] See also Fig. 9 As shown, the structured block diagram of the electronic device 900 that can be used as the server or client of the present application will now be described, which is an example of the hardware device that can be applied to various aspects of the present application. The electronic device is intended to represent the computer device of various forms of digital electronics, such as, laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, personal digital processing, cellular phones, smart phones, wearable devices and other similar computing devices. The components shown herein, their connections and relationships, and their functions are only used as examples, and are not intended to limit the implementation of the present application described herein and / or required.

[0180] like Fig. 9As shown, the electronic device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 to a random access memory (RAM) 903. In the RAM 903, various programs and data required for the operation of the device 900 can also be stored. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0181] A plurality of components in the electronic device 900 are connected to the I / O interface 905, including: an input unit 906, an output unit 907, a storage unit 908, and a communication unit 909. The input unit 906 may be any type of device capable of inputting information to the electronic device 900, and the input unit 906 may receive input digital or character information, and generate key signal inputs related to user settings and / or function control of the electronic device. The output unit 907 may be any type of device capable of presenting information, and may include but is not limited to a display, a speaker, a video / audio output terminal, a vibrator, and / or a printer. The storage unit 908 may include but is not limited to a disk, an optical disk. The communication unit 909 allows the electronic device 900 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks, and may include but is not limited to a modem, a network card, an infrared communication device, a wireless communication transceiver, and / or a chipset, such as a Bluetooth device, a WiFi device, a worldwide interoperability for microwave access (WiMax) device, a cellular communication device, and / or the like.

[0182] The computing unit 901 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 901 performs the various methods and processes described above. For example, in some embodiments, the training method of the above-mentioned business recommendation model may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 908. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 900 via the ROM 902 and / or the communication unit 909. In some embodiments, the computing unit 901 may be configured to perform the training method of the above-mentioned business recommendation model by any other appropriate means (e.g., by means of firmware).

[0183] The program code for implementing the method of the present application can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, implements the functions / operations specified in the flow chart and / or block diagram. The program code can be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0184] In the context of the present application, 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 RAM, a ROM, an erasable programmable read-only memory (EPROM) or a flash memory, an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0185] As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0186] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a cathode ray tub (CRT) or a liquid crystal display (LCD) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0187] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by 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), and the Internet.

[0188] A computer system may include clients and servers. Clients and servers are generally remote from each other and usually interact through a communication network. The relationship of client and server is generated by computer programs running on respective computers and having a client-server relationship to each other.

[0189] Furthermore, it should be understood that what is disclosed above is only a preferred embodiment of the present application, and certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope covered by the present application.

Claims

1. A training method for a business recommendation model, characterized in that: include: Obtain a training sample set for a business recommendation model; The training sample set includes: a first sample set for recommending a target service to a user and a second sample set for not recommending the target service to the user; Based on the gradient information obtained by training the service recommendation model with the training sample set, the network parameters corresponding to the multiple networks included in the service recommendation model are updated; wherein the multiple networks include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network for generating the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the service usage rate after the target service is recommended to the user, and the second network is used to determine the service usage rate after the target service is not recommended to the user.

2. The method according to claim 1, characterized in that The step of obtaining a training sample set for a business recommendation model includes: Obtaining user information corresponding to a plurality of users respectively; each piece of user information includes a plurality of user features of the corresponding user; Based on a preset feature processing rule, perform feature dimensionality reduction processing on the plurality of user information to obtain the plurality of user information after the feature dimensionality reduction processing; The plurality of user information after the feature dimension reduction processing is used as the training sample set.

3. The method according to claim 1, characterized in that The updating of network parameters corresponding to a plurality of networks included in the business recommendation model respectively based on the gradient information obtained by training the business recommendation model based on the training sample set comprises: Based on the label loss values ​​respectively corresponding to the multiple training samples included in the training sample set, determining the gradient information respectively corresponding to the shared network and the super network included in the multiple networks; wherein the shared network is used to perform shared feature extraction on the first sample set and the second sample set; The network parameters of the shared network are updated based on the gradient information corresponding to the shared network, and the network parameters of the super network are updated based on the gradient information corresponding to the super network.

4. The method according to claim 3, characterized in that Before the label loss values ​​corresponding to the multiple training samples included in the training sample set, the method further includes: For the multiple training samples, the following operations are performed respectively: Extracting shared features from a training sample through the shared network to obtain the shared features of the training sample; Inputting the shared feature into the first network and the second network respectively, obtaining a first service usage rate output by the first network and a second service usage rate output by the second network; Determining a prediction label of the one training sample based on the first service usage rate and the second service usage rate; A label loss value is obtained based on the predicted label and the true label of the one training sample.

5. The method according to claim 4, characterized in that The determining, based on the first service usage rate and the second service usage rate, a predicted label of the one training sample includes: Based on the first service usage rate and the second service usage rate, obtaining a corresponding service usage rate difference; When the service usage rate difference is greater than a set usage rate difference threshold, the predicted label of the one training sample is determined to be a target user.

6. The method according to claim 3, 4 or 5, characterized in that The updating of the network parameters of the super network based on the gradient information corresponding to the super network includes: Based on a preset neuron discarding algorithm, determining a plurality of neurons in the hypernetwork that participate in the current model training; The network parameters of the plurality of neurons are updated based on the gradient information corresponding to the super network.

7. The method according to any one of claims 1 to 5, characterized in that The output layer of the hypernetwork includes a plurality of neuron groups, and each neuron group is used to output part of the plurality of parameters output by the hypernetwork.

8. A training device for a business recommendation model, characterized in that: include: A sample acquisition module is used to obtain a training sample set for a business recommendation model; The training sample set includes: a first sample set for recommending a target service to a user and a second sample set for not recommending the target service to the user; A model training module, used to update the network parameters corresponding to the multiple networks included in the business recommendation model based on the gradient information obtained by training the business recommendation model with the training sample set; wherein the multiple networks include at least: a first network set for the first sample set and a second network set for the second sample set, and a super network that generates the network parameters corresponding to the first network and the second network, respectively, the first network is used to determine the business usage rate after recommending the target business to the user, and the second network is used to determine the business usage rate after not recommending the target business to the user.

9. An electronic device, comprising: processor; as well as Memory for storing programs, The program includes instructions, which, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1 to 7.

Images