Multi-objective recommendation method, device, computing device and medium

Abstract

The disclosed embodiments provide a multi-objective recommendation method, apparatus, computing device and medium. The method includes: extracting input features from the user information of the user and the object information of the object to be recommended; based on the input features, determining the target score of each business target among multiple business targets of the object to be recommended through a multi-objective recommendation model; weighting the target score of each business target based on the weight corresponding to each business target, and determining the recommendation score of the object to be recommended based on the result of the weighted processing, the weight is determined according to the expected score of each business target of the object to be recommended in the current business scenario; recommending the object to be recommended to the user based on the recommendation score of the object to be recommended. According to the technical solution of the disclosed embodiments, multiple business targets can be optimized at the same time, and the benefits of each business target can be balanced and adjusted.

Description

Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] With the development of Internet technology, the amount of information is growing exponentially. As a technical means to solve information overload and tap into users' potential needs, recommendation plays an important role in many fields, such as e-commerce, news information, and movie recommendations.

[0004] In the relevant technical solution, a recommendation ranking model based on deep learning is adopted, and the user's historical behavior data is used as a sample to train the model. In the prediction stage, the recalled objects to be recommended, such as goods and information, are scored and ranked, and the top N recommended objects are selected for display to complete the recommendation process. Summary of the invention

[0005] However, the recommendation ranking model in the above technical solution generally only uses click-through rate as the optimization business goal. In real business scenarios, it is often necessary to optimize multiple business goals, such as the dwell time indicator for information and the click-to-purchase rate indicator for e-commerce. The above technical solution is difficult to optimize multiple business goals and cannot balance the benefits of each business goal.

[0006] Therefore, an improved multi-objective recommendation method is needed to optimize multiple business objectives and balance the benefits of each business objective.

[0007] In a first aspect of an embodiment of the present disclosure, a multi-objective recommendation method is provided, comprising: extracting input features from user information of a user and object information of an object to be recommended, the input features comprising: user features and object features; based on the input features, determining a target score for each of multiple business targets of the object to be recommended through a multi-objective recommendation model, the multi-objective recommendation model being a neural network model for determining the target scores of each business target of the object to be recommended; weighting the target scores of each of the business targets based on weights corresponding to each of the business targets, and determining a recommendation score for the object to be recommended based on a result of the weighted processing, the weight being determined according to the expected scores of each of the business targets of the object to be recommended in the current business scenario; and recommending the object to be recommended to the user based on the recommendation score of the object to be recommended.

[0008] In some example embodiments of the present disclosure, the method further includes: initializing the weights corresponding to each of the business objectives; adjusting the weights corresponding to each of the business objectives in a step-by-step manner, and obtaining a set of weights after each adjustment; determining a recommendation score corresponding to each set of weights after the weight adjustment; determining a difference between the recommendation score and an expected recommendation score, wherein the expected recommendation score is an expected score of the object to be recommended in the current business scenario; and using at least one set of weights corresponding to the recommendation score whose difference is less than a predetermined threshold as the weight of each of the business objectives of the object to be recommended.

[0009] In some example embodiments of the present disclosure, the input layer of the multi-objective recommendation model includes a first network layer and a second network layer, the input features further include cross features, and extracting the input features from the user information of the user and the object information of the object to be recommended includes: obtaining original features from the user information of the user and the object information of the object to be recommended through the first network layer, the original features including the user features and the object features; obtaining the cross features from the user information of the user and the object information of the object to be recommended through the second network layer; and concatenating the original features and the cross features to generate the input features.

[0010] In some example embodiments of the present disclosure, the first network layer is a multi-layer perceptron network, the second network layer is an inner product-based neural network IPNN layer, and the obtaining of cross features from the user information of the user and the object information of the object to be recommended through the second network layer includes: obtaining user features and object features from the user information of the user and the object information of the object to be recommended through the IPNN layer; performing inner product processing on the user features and / or the object features to generate cross features.

[0011] In some example embodiments of the present disclosure, the hidden layer of the multi-objective recommendation model includes multiple sub-networks and multiple gating layers, and determining the target score of each of the multiple business objectives for the object to be recommended through the multi-objective recommendation model includes: determining the sub-features corresponding to the sub-networks in the input features through each of the sub-networks; determining the weights of each of the sub-features corresponding to the business objectives through the gating layers corresponding to the business objectives; performing weighted processing on each of the sub-features based on the weights of the sub-features to obtain target features corresponding to the business objectives; and determining the target scores of each business objective based on the target features corresponding to each of the business objectives.

[0012] In some example embodiments of the present disclosure, the subnetwork includes a first cross-feature extraction network and a second cross-feature extraction network, and the sub-features corresponding to the subnetwork in the input features are determined through each of the subnetworks, including: extracting high-order cross-features corresponding to the subnetwork in the input features through the first cross-feature extraction network; extracting second-order cross-features corresponding to the subnetwork in the input features through the second cross-feature extraction network, wherein the order of the high-order cross-features is greater than the order of the second-order cross-features; and concatenating the high-order cross-features and the second-order cross-features to generate sub-features corresponding to the subnetwork.

[0013] In some example embodiments of the present disclosure, the first cross feature extraction network is a multi-layer perceptron network, and the second cross feature extraction network is a factorization machine network.

[0014] In some example embodiments of the present disclosure, the method further includes: acquiring sample data, the sample data including sample features and sample labels, the sample features including user features of sample users and object features of sample objects, the sample labels including actual target scores for each business target of the sample objects; inputting the sample data into the multi-objective recommendation model to determine predicted target scores for each business target of the sample objects; determining a loss function corresponding to the business target based on a difference between the actual target score and the predicted target score; and adjusting parameters of the multi-objective recommendation model based on the loss function.

[0015] In some example embodiments of the present disclosure, determining the recommendation score of the object to be recommended based on the result of weighted processing includes: determining the recommendation score of the object to be recommended by a Sigmoid activation function based on the result of weighted processing.

[0016] In some example embodiments of the present disclosure, the multi-objective recommendation model is a neural network model based on soft parameter sharing.

[0017] In a second aspect of the embodiments of the present disclosure, a multi-objective recommendation device is provided, comprising: an input feature extraction module, used to extract input features from user information of a user and object information of an object to be recommended, the input features comprising: user features and object features; a target score determination module, used to determine, based on the input features, a target score for each of multiple business objectives of the object to be recommended through a multi-objective recommendation model, the multi-objective recommendation model being a neural network model for determining the target score of each business objective of the object to be recommended; a recommendation score determination module, used to perform weighted processing on the target scores of each of the business objectives based on the weights corresponding to each of the business objectives, and determine the recommendation score of the object to be recommended based on the result of the weighted processing, the weight being determined according to the expected scores of each of the business objectives of the object to be recommended in the current business scenario; a recommendation module, used to recommend the object to be recommended to the user based on the recommendation score of the object to be recommended.

[0018] In some example embodiments of the present disclosure, the device also includes: a weight initialization module, which initializes the weights corresponding to each of the business objectives; a weight adjustment module, which adjusts the weights corresponding to each of the business objectives in a step-by-step manner, and obtains a set of weights after each adjustment; an adjustment score determination module, which determines the recommendation score corresponding to each set of weights after the weight adjustment; a difference determination module, which determines the difference between the recommendation score and the expected recommendation score, wherein the expected recommendation score is the expected score of the object to be recommended in the current business scenario; and a weight selection module, which uses at least one set of weights corresponding to the recommendation score whose difference is less than a predetermined threshold as the weight of each of the business objectives of the object to be recommended.

[0019] In some example embodiments of the present disclosure, the input layer of the multi-objective recommendation model includes a first network layer and a second network layer, the input features further include cross features, and the input feature extraction module includes: an original feature extraction unit, used to obtain original features from the user information of the user and the object information of the object to be recommended through the first network layer, the original features including the user features and the object features; a cross feature extraction unit, used to obtain the cross features from the user information of the user and the object information of the object to be recommended through the second network layer; a first splicing processing unit, used to splice the original features and the cross features to generate the input features.

[0020] In some example embodiments of the present disclosure, the first network layer is a multi-layer perceptron network, the second network layer is an inner product-based neural network IPNN layer, and the cross feature extraction unit is further used to: obtain user features and object features from the user information of the user and the object information of the object to be recommended through the IPNN layer; perform inner product processing on the user features and / or the object features to generate cross features.

[0021] In some example embodiments of the present disclosure, the hidden layer of the multi-objective recommendation model includes multiple sub-networks and multiple gating layers, and the target score determination module includes: a sub-feature determination module, used to determine the sub-features corresponding to the sub-networks in the input features through each of the sub-networks; a weight determination unit, used to determine the weights of each of the sub-features corresponding to the business goal through the gating layer corresponding to the business goal; a weighted processing unit, used to perform weighted processing on each of the sub-features based on the weights of the sub-features to obtain target features corresponding to the business goals; and a score determination unit, used to determine the target scores of each business goal based on the target features corresponding to each of the business goals.

[0022] In some example embodiments of the present disclosure, the sub-network includes a first cross-feature extraction network and a second cross-feature extraction network, and the sub-feature determination module is further used to: extract high-order cross-features corresponding to the sub-network from the input features through the first cross-feature extraction network; extract second-order cross-features corresponding to the sub-network from the input features through the second cross-feature extraction network, wherein the order of the high-order cross-features is greater than the order of the second-order cross-features; and concatenate the high-order cross-features and the second-order cross-features to generate sub-features corresponding to the sub-network.

[0023] In some example embodiments of the present disclosure, the first cross feature extraction network is a multi-layer perceptron network, and the second cross feature extraction network is a factorization machine network.

[0024] In some example embodiments of the present disclosure, the device also includes: a sample acquisition module, used to acquire sample data, the sample data including sample features and sample labels, the sample features including user features of sample users and object features of sample objects, and the sample labels including actual target scores for each business target of the sample objects; a predicted score determination module, used to input the sample data into the multi-objective recommendation model to determine the predicted target scores for each business target of the sample objects; a loss function determination module, used to determine the loss function corresponding to the business target based on the difference between the actual target score and the predicted target score; and a parameter adjustment module, used to adjust the parameters of the multi-objective recommendation model based on the loss function.

[0025] In some example embodiments of the present disclosure, the recommendation score determination module is further used to: determine the recommendation score of the object to be recommended by using a Sigmoid activation function based on a result of weighted processing.

[0026] In some example embodiments of the present disclosure, the multi-objective recommendation model is a neural network model based on soft parameter sharing.

[0027] In a third aspect of an embodiment of the present disclosure, a computing device is provided, including: a processor and a memory, the memory storing executable instructions, the processor being used to call the executable instructions stored in the memory to execute a method as described in any one of the above-mentioned first aspects.

[0028] In a fourth aspect of the embodiments of the present disclosure, a medium is provided, on which a program is stored, and when the program is executed by a processor, the method as described in any one of the first aspects above is implemented.

[0029] According to the technical solution of the embodiment of the present disclosure, on the one hand, the target score for each business goal of the object to be recommended is determined by a multi-objective recommendation model, so that multiple business goals of the object to be recommended can be optimized at the same time; on the other hand, the weight of each business goal is determined according to the expected score of each business goal of the object to be recommended in the current business scenario, and the weight of each business goal can be adjusted according to business needs, so that the benefits of each business goal can be balanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The above and other purposes, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0031] Figure 1 A block diagram schematically shows an application scenario of the multi-objective recommendation method according to an embodiment of the present disclosure;

[0032] Figure 2 The flowchart of the multi-objective recommendation method according to some embodiments of the present disclosure is schematically shown;

[0033] Figure 3 A block diagram schematically shows a recommendation system to which a multi-objective recommendation method according to some embodiments of the present disclosure is applied;

[0034] Figure 4 Schematically shows a flow chart of weight adjustment according to some embodiments of the present disclosure;

[0035] Figure 5A schematic diagram schematically illustrates a computer-readable storage medium according to some embodiments of the present disclosure;

[0036] Figure 6 The structure block diagram of the multi-objective recommendation device according to some embodiments of the present disclosure is schematically shown;

[0037] Figure 7 The structure block diagram of a computing device according to some embodiments of the present disclosure is schematically shown.

[0038] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0039] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0040] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0041] According to an embodiment of the present disclosure, a multi-objective recommendation method, apparatus, computing device and medium are proposed.

[0042] In this article, it is important to understand that the terms involved are:

[0043] Object to be recommended: indicates the object recommended to the user in the current business scenario. For example, in the shopping business scenario, the object to be recommended is the product; in the news information scenario, the object to be recommended is the information.

[0044] Business goal: a business indicator used to determine whether to recommend an object to be recommended to a user. For example, a business goal may be a click-through rate indicator, a dwell time indicator, a conversion rate indicator, etc.

[0045] Subnetwork layer: also known as expert network layer, a neural network layer used to characterize at least one type of input features in the current business scenario. For example, suppose the subnetwork layer of the multi-objective recommendation model contains three subnetworks. The first subnetwork can be used to characterize the features corresponding to user browsing operations, the second subnetwork can be used to characterize the features corresponding to purchase operations, and the third subnetwork can be used to characterize the features corresponding to collection operations.

[0046] Gating layer: used to determine the weights of the output features of each sub-network corresponding to the business goal.

[0047] Multi-objective recommendation model: a model used to determine the target scores of various business objectives of the recommended object. The multi-objective recommendation model can be a neural network model based on soft parameter sharing, such as an MMOE (Multi-gate Mixture of Experts) model.

[0048] IPNN (Inner Product-based Neural Network) is a DNN (Deep Neural Networks) network structure based on the idea of ​​product layer and uses multiplication operation to reflect feature crossover, that is, the inner product method is used for each feature to obtain crossover features.

[0049] FM (Factorization Machines): Factorization machine is a machine learning algorithm based on matrix decomposition. The specific principle is as follows:

[0050] The common linear model in machine learning is shown in the following formula (1):

[0051] y=w 1 x 1 +w 2 x 2 +…+w n x n (1)

[0052] The disadvantage of linear models is that they cannot capture the relationship between features. To solve this problem, polynomial regression is introduced as follows (2):

[0053]

[0054] In formula (2), since the weight w between every two features needs to be learned, n(n-1) / 2 parameters need to be learned, where n is the number of features. The value of n is generally large, which will make the model particularly complex. Since the data in reality is often sparse, the model will not be able to learn the parameters. Inspired by matrix decomposition, factor decomposition machines are introduced to solve the above problem.

[0055] The factorization machine FM is based on the idea of ​​matrix decomposition and decomposes the sparse matrix into the product of two low-order matrices, namely w i,j =v i v j , the above formula (2) can be transformed into the following formula (3):

[0056]

[0057] In the above formula (2), the w parameter cannot be learned when the data is sparse, while the w parameter in formula (3) can be approximated by multiplying the latent vectors of i and j, thus solving the problem that the parameters cannot be learned with sparse data.

[0058] In addition, any number of elements in the drawings is for illustration and not limitation, and any naming is only for distinction and does not have any limiting meaning.

[0059] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0061] The inventors have found that in order to solve the problem of optimizing multiple objectives simultaneously, in a technical solution, a multi-task model with a shared underlying layer is used to determine the target scores of each business objective of the recommended object. However, in this technical solution, conflicts often occur between the various optimized business objectives, making it difficult to balance the benefits of each business objective.

[0062] Based on the above content, the basic idea of ​​the present disclosure is to determine the target score of each business target of the object to be recommended through a multi-target recommendation model, perform weighted processing on the target score of each business target based on the weight corresponding to each business target, and determine the weight corresponding to each business target according to the expected score of each business target of the object to be recommended in the current business scenario. The recommendation score of the object to be recommended is determined based on the result of the weighted processing, and the object to be recommended is recommended to the user based on the recommendation score. According to the technical solution of the embodiment of the present disclosure, on the one hand, the target score for each business target of the object to be recommended is determined through a multi-target recommendation model, so that multiple business targets of the object to be recommended can be optimized at the same time; on the other hand, the weight of each business target is determined according to the expected score of each business target of the object to be recommended in the current business scenario, and the weight of each business target can be adjusted according to business needs, so that the benefits of each business target can be balanced.

[0063] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0064] Application Scenario Overview

[0065] It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0066] Figure 1 A block diagram schematically shows an application scenario of the multi-objective recommendation method according to an embodiment of the present disclosure.

[0067] Reference Figure 1 As shown, the application scenario may include: at least one client 110 and a server 120, wherein the client 110 is installed with multiple applications, such as shopping applications, video applications, news information applications, etc. The client 110 and the server 120 communicate with each other through a network 130. Taking the shopping application scenario as an example, the business objectives are click-through rate and purchase rate. The user opens the shopping application installed on the client 110 and enters the name of the product to be searched in the search box. The server 120 determines the target scores of the click-through rate and purchase rate for the product to be recommended through a multi-objective recommendation model based on the user information of the user and the product information of the product to be recommended, and performs weighted processing on the target scores according to the weights corresponding to the click-through rate and the purchase rate to obtain the recommendation score of the product to be recommended, and recommends the product to the user based on the recommendation score.

[0068] It should be noted that the client 110 may be a mobile phone, a tablet computer, a desktop computer, a portable notebook computer, or a vehicle-mounted terminal, etc. The server 120 may be a physical server including an independent host, or a virtual server carried by a host cluster, or a cloud server. The network 130 may be a wired network or a wireless network, for example, the network 130 may be a PSTN (Public Switched Telephone Network) or the Internet.

[0069] Exemplary Methods

[0070] In combination with the above application scenarios, refer to Figure 2 To describe the multi-objective recommendation method according to an exemplary embodiment of the present disclosure. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0071] Reference Figure 2 As shown, in step S210, input features are extracted from the user information of the user and the object information of the object to be recommended, and the input features include: user features and object features.

[0072] In the example embodiment, taking the shopping scenario as an example, the user information of the user may include basic user information and historical user behavior information. The basic user information may include the user's age, gender, occupation, etc. The historical user behavior information may include the user's historical purchase information, collection information, comment information, etc. The object to be recommended is a product, and the object information of the object to be recommended may include information such as product name, product model, product size, and product color.

[0073] Further, in an example embodiment, input features may be extracted from the user information of the user and the product information of the product through a feature extraction neural network, for example, input features may be extracted from the user information of the user and the product information of the product through a multi-layer perceptron neural network. The input features may include user features and product features.

[0074] In step S220, based on the input features, a target score for each of the multiple business objectives of the object to be recommended is determined by a multi-objective recommendation model, where the multi-objective recommendation model is a neural network model that determines the target score for each business objective of the object to be recommended.

[0075] In an example embodiment, the multi-objective recommendation model may be a neural network model based on soft parameter sharing, for example, the multi-objective recommendation model may be an MMOE model. The business goal is a business goal in the current business scenario. Taking the business scenario as a shopping scenario as an example, the business goal may be the click-through rate of the product, the click-to-purchase rate, the page dwell time, etc.

[0076] Further, the input features extracted in step S210 are input into the multi-objective recommendation model, and the target scores of the various business objectives for the to-be-recommended products are determined by the multi-objective recommendation model. In an exemplary embodiment, the multi-objective recommendation model includes multiple sub-networks and multiple gating layers, and determining the target scores of the various business objectives for the to-be-recommended objects by the multi-objective recommendation model includes: inputting the input features into the sub-networks; determining the sub-features corresponding to the sub-networks in the input features through the various sub-networks; determining the weights of the various sub-features corresponding to the business objectives through the gating layers corresponding to the business objectives; performing weighted processing on the various sub-features based on the weights of the sub-features to obtain the target features corresponding to the business objectives; and determining the target scores of the various business objectives based on the target features corresponding to the various business objectives.

[0077] Among them, the subnetwork is a neural network layer used to characterize at least one type of input features in the current business scenario. For example, the subnetwork layer of the multi-objective recommendation model contains three subnetworks, the first subnetwork is used to characterize the features corresponding to the user's browsing operation, the second subnetwork is used to characterize the features corresponding to the purchase operation, and the third subnetwork is used to characterize the features corresponding to the collection operation. The gating layer corresponds to the business goal in the current business scenario, and is used to determine the weights of the output features of each subnetwork corresponding to the business goal, and weightedly process the output features of each expert network to obtain the target features corresponding to the business goal.

[0078] Furthermore, the target features corresponding to the business target may be input into the target score prediction network to determine the target score corresponding to the business target. In an example embodiment, the target score prediction network may be a tower network in the MMOE model.

[0079] In step S230, the target scores of each business target are weighted based on the weights corresponding to each business target, and the recommendation score of the object to be recommended is determined based on the result of the weighted processing, and the weight is determined according to the expected scores of each business target of the object to be recommended in the current business scenario.

[0080] In an example embodiment, the expected scores of various business objectives in the current business scenario can be set in advance, and the weights corresponding to various business objectives can be determined based on the sizes of the expected scores of various business objectives in the current business scenario. For example, for shopping scenarios, since the click-through purchase rate is more important, that is, the expected score of the click-through purchase rate is higher, the weight corresponding to the click-through purchase rate can be set higher; for news and information, since the page dwell time is more important, that is, the expected score of the page dwell time is higher, the weight of the page dwell time can be set higher.

[0081] Furthermore, the target scores of each business target in the multiple business targets are weighted based on the weights corresponding to each business target to determine the recommendation score of the object to be recommended. For example, the recommendation score of the object to be recommended can be determined based on the result of the weighted processing by using the Sigmoid activation function.

[0082] In step S240, the object to be recommended is recommended to the user based on the recommendation score of the object to be recommended.

[0083] In an exemplary embodiment, the commodities to be recommended are sorted based on their recommendation scores, and the commodities to be recommended whose recommendation scores are greater than a predetermined threshold are selected and recommended to the user.

[0084] according to Figure 2 The technical solution in the example embodiment, on the one hand, determines the target score of each business goal of the recommended object through a multi-objective recommendation model, and can optimize multiple business goals of the recommended object at the same time; on the other hand, determines the weight of each business goal according to the expected score of each business goal of the recommended object in the current business scenario, and can adjust the weight of each business goal according to business needs, so as to balance the benefits of each business goal.

[0085] In addition, before applying the multi-objective recommendation model, the multi-objective recommendation model needs to be trained. Therefore, in an example embodiment, the multi-objective recommendation method further includes: obtaining sample data, the sample data includes sample features and sample labels, the sample features include user features of sample users and object features of sample objects, and the sample labels include actual target scores for each business target of the sample objects; inputting the sample data into the multi-objective recommendation model to determine the predicted target scores for each business target of the sample objects; determining the loss function corresponding to the business target based on the difference between the actual target score and the predicted target score; and adjusting the parameters of the multi-objective recommendation model based on the loss function.

[0086] It should be noted that different loss functions can be set for different business goals. For example, business goals can be divided into classification business goals and regression business goals. For example, determining whether a product link will be clicked is a classification business goal, and the page dwell time is a regression business goal. The cross entropy loss function is used for the classification business goal, and the square loss function is used for the regression business goal.

[0087] Figure 3 The block diagram of a recommendation system to which the multi-objective recommendation method according to some embodiments of the present disclosure is schematically shown.

[0088] Reference Figure 3 As shown, the user log module 305 is used to obtain the user's historical behavior data. Taking the shopping scenario as an example, the user's historical behavior data may include purchased product data, favorite product data, browsed product data, commented product data, etc.

[0089] The sample module 310 is used to generate sample data based on the user's historical behavior data. The sample data may include the user's user information, the object information of the sample object, and the sample label. Taking the business goal as the click-through rate target as an example, the sample label of the user clicking the product page link after the product is exposed is a positive sample, and the sample label of the user who is exposed but not clicked is a negative sample.

[0090] The input feature extraction module 315 includes a multi-layer perceptron 312 and an IPNN (Inner Product-based Neural Network) 314. The multi-layer perceptron 312 is used to obtain original features from the user information of the user and the object information of the object to be recommended, and the original features include user features and object features; the IPNN 314 is used to obtain cross features from the user information of the user and the object information of the object to be recommended. The input feature extraction module 315 is used to splice the original features extracted by the multi-layer perceptron 312 and the cross features extracted by the IPNN 314 to generate input features.

[0091] In the exemplary embodiment, important features in the sample features, such as user features, commodity category features, etc., can be used to generate cross features using the network structure of IPNN. i represents the embedding vector of the i-th feature, e j represents the embedding vector of the jth feature, then e is generated based on the inner product i With e j The cross-features can be expressed as the following formula (4):

[0092]

[0093] Among them, l p is the crossover feature, is the weight matrix, p={p i,j}, i = 1…N, j = 1…N, p i,j = <e i ,e j >.

[0094] Furthermore, let the original feature of the sample data be l z , then the input feature extraction module 315 extracts the original feature l by the following formula (5): z With cross feature l p Perform concatenation to obtain the input feature z:

[0095] z=concat([l z ,l p ]) (5)

[0096] By extracting the cross features of the sample data and concatenating the original features of the sample data with the cross features, richer model features of the sample data can be obtained, making the prediction results of the trained multi-objective recommendation model more accurate.

[0097] Continue to refer to Figure 3 As shown, the sub-network layer includes n sub-networks 320, and each sub-network 320 includes a multi-layer perceptron 322 and a factor decomposition machine 324. The multi-layer perceptron 322 is used to extract high-order cross features corresponding to the sub-network in the input features; the factor decomposition machine 324 is used to extract second-order cross features corresponding to the sub-network 320 in the input features. The sub-network layer is used to concatenate the high-order cross features extracted by the multi-layer perceptron 322 and the second-order cross features extracted by the factor decomposition machine 324 to generate sub-features corresponding to the sub-network 320.

[0098] For example, suppose the vector output by the multilayer perceptron 322 is y dnn , the vector output by the factorization machine 324 is y FM , y dnn With y FMThe concatenation process is performed by the following formula, and the result after the concatenation process is used as the sub-feature output by the sub-network 320, as shown in the following formula (6):

[0099] y=concat([y FM ,y dnn ]) (6)

[0100] Among them, y is the sub-feature after splicing, y FM is the vector output by the factorization machine 324, y dnn is the vector output by the multilayer perceptron 322.

[0101] By combining the high-order feature extraction network with the low-order feature extraction network, the multi-target recommendation model can more effectively learn the cross-information of high-order features and low-order features, making the model prediction results more accurate.

[0102] Next, the gating layer 325 is used to determine the weights of the sub-features output by each sub-network 320 corresponding to the business goal; each sub-feature is weighted based on the weight of the sub-feature to obtain the target feature corresponding to the business goal. For example, suppose the shared bottom layer of the multi-objective recommendation model is divided into n different sub-networks, namely, the expert network, and each sub-network is fused with the kth gating layer. The fusion result is expressed by the following formula (7):

[0103]

[0104] Among them, f k (x) is the fusion result of the sub-network and the k-th gating layer, f i (x) is the sub-feature output by the ith sub-network, g k (x) i is the weight corresponding to the sub-feature of the ith sub-network output by the kth gating layer.

[0105] Furthermore, the kth gating layer can be expressed by the following formula (8):

[0106] g k (x) = soft max(W gk x) (8)

[0107] Among them, W gk is the trained weight matrix, and x is the input of the gating layer.

[0108] In an example embodiment, the business goal network layer may include multiple business goal networks, the input of the business goal network is the target feature output by the gating layer 325, and the output is the target score of the business goal. Figure 3In the example, the service target network layer includes service target 1 network 326, service target 2 network 328 and service target 3 network 330. The output of service target k network can be expressed by the following formula (9):

[0109] y k =h k (f k (x)) (9)

[0110] Among them, h k represents the kth service target network, i.e., the tower network, y k Represents the output of the kth business target network.

[0111] By adding a gating layer to the multi-objective recommendation model, different sample data can use the shared layer, i.e., the sub-network layer, in a diversified manner, thereby weakening the impact of the differences in the relevance of different targets. At the same time, the transfer learning method can be used to enable each business target to learn information from each other.

[0112] Continue to refer to Figure 3 As shown, the score fusion network 335 fuses the target scores output by each business target network. For example, suppose the output results of k business target networks are logit i , where i=1, 2, ..., k, the recommendation score output by the score fusion network 335 is as follows (10):

[0113]

[0114] in, is the weight corresponding to each business objective, and Output is the final recommendation score output by the score fusion network 335. The larger the value is, the greater the impact of the business goal on the overall output recommendation score is, and the higher the corresponding online indicator is.

[0115] The offline test module 340 is used to adjust the weights corresponding to the various business objectives in a step-by-step manner according to the expected scores of the various business objectives in the current business scenario under offline test conditions, and select the weight parameters with better offline results to go online. For example, the offline test module 340 can adjust the weights corresponding to the various business objectives in an equidistant step manner, or in a non-equidistant step manner.

[0116] The sorting module 345 is used to sort the objects to be recommended according to their recommendation scores, and select the objects to be recommended whose recommendation scores are greater than a predetermined threshold to recommend to the user.

[0117] Figure 4The following schematically shows a flowchart of weight adjustment according to some embodiments of the present disclosure.

[0118] Reference Figure 4 As shown, in step S410, the weights corresponding to the various business objectives are initialized.

[0119] In an example embodiment, the weight corresponding to each business goal may be initialized to 0, or may be initialized to other appropriate empirical values. For example, taking a shopping scenario as an example, business goals may include: click-through rate goal, purchase rate goal, and add-to-cart rate goal, and the corresponding set of weights is {0, 0, 0}.

[0120] In step S420, the weights corresponding to the various business objectives are adjusted in a step-by-step manner, and a set of weights is obtained after each adjustment.

[0121] In the example embodiment, the weights corresponding to the various business objectives may be adjusted in an equidistant step manner, or in a non-equidistant step manner. It is also possible to adjust the weights of one business objective in an equidistant step manner, and adjust the weights of another business objective in a non-equidistant step manner.

[0122] In step S430 , a recommendation score corresponding to each set of weights after weight adjustment is determined.

[0123] In an example embodiment, after weight adjustment, a set of weights corresponding to multiple business objectives is obtained, the target scores of each business objective are weighted based on the weights corresponding to each business objective, and the recommendation score corresponding to the set of weights is determined based on the weighted processing structure.

[0124] In step S440, the difference between the recommendation score and the expected recommendation score is determined, where the expected recommendation score is the expected score of the object to be recommended in the current business scenario.

[0125] In an example embodiment, an expected score of the object to be recommended, ie, an expected recommendation score, may be set for the current business scenario, and a difference between the recommendation score corresponding to each group of weights and the expected recommendation score may be determined.

[0126] In step S450, at least one set of weights corresponding to the recommendation scores whose difference is less than a predetermined threshold is used as the weights of each business goal of the object to be recommended.

[0127] In an example embodiment, at least one set of weights corresponding to the recommendation scores whose difference is less than a predetermined threshold is determined, and the at least one set of weights is used as the weights of each business goal of the object to be recommended. The predetermined threshold can be determined according to the number of objects to be recommended. For example, if there are multiple recommendation scores whose difference is less than the predetermined threshold, the set of weights with the smallest difference can be selected as the weights of each business goal of the object to be recommended, or multiple sets of weights corresponding to multiple recommendation scores can be respectively used as the weights of each business goal of the object to be recommended.

[0128] according to Figure 4 The technical solution in the example embodiment adjusts the weights corresponding to each business goal in a step-by-step manner, and can adjust the weights of each business goal in real time according to the needs of the business goal, thereby dynamically balancing the benefits of each business goal.

[0129] Exemplary Media

[0130] After introducing the method of the exemplary embodiment of the present disclosure, next, the medium of the exemplary embodiment of the present disclosure will be described.

[0131] In some possible embodiments, various aspects of the present disclosure may also be implemented as a medium having program code stored thereon, which, when executed by a processor of a device, is used to implement the steps of the multi-objective recommendation method according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification.

[0132] In some possible embodiments, the processor of the device is used to implement the following steps when executing the program code: extracting input features from the user information of the user and the object information of the object to be recommended, the input features including: user features and object features; based on the input features, determining the target score of each business goal among multiple business goals of the object to be recommended through a multi-objective recommendation model, the multi-objective recommendation model is a neural network model for determining the target score of each business goal of the object to be recommended; weighting the target score of each business goal based on the weight corresponding to each business goal, and determining the recommendation score of the object to be recommended based on the result of the weighted processing, the weight being determined according to the expected score of each business goal of the object to be recommended in the current business scenario; and recommending the object to be recommended to the user based on the recommendation score of the object to be recommended.

[0133] refer to Figure 5 As shown, a program product 500 for implementing the above data processing method according to an embodiment of the present disclosure is described, which can adopt a portable compact disk read-only memory and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0134] It should be noted that the above-mentioned medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory, a read-only memory, an erasable programmable read-only memory, an optical fiber, a portable compact disk read-only memory, an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0135] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, wherein readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to: electromagnetic signals, optical signals, or any suitable combination of the above. A readable signal medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0136] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, radio frequency signals, etc., or any suitable combination of the foregoing.

[0137] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network or a wide area network, or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0138] Exemplary Devices

[0139] After introducing the medium of the exemplary embodiment of the present disclosure, next, reference is made to Figure 6 A multi-objective recommendation apparatus according to an exemplary embodiment of the present disclosure is described.

[0140] Reference Figure 6As shown, the multi-objective recommendation device 600 includes: an input feature extraction module 610, which is used to extract input features from the user information of the user and the object information of the object to be recommended, and the input features include: user features and object features; a target score determination module 615, which is used to determine the target score of each business target among multiple business targets of the object to be recommended through a multi-objective recommendation model based on the input features, and the multi-objective recommendation model is a neural network model for determining the target score of each business target of the object to be recommended; a recommendation score determination module 620, which is used to weight the target score of each business target based on the weight corresponding to each business target, and determine the recommendation score of the object to be recommended based on the result of the weighted processing, and the weight is determined according to the expected score of each business target of the object to be recommended in the current business scenario; a recommendation module 625, which is used to recommend the object to be recommended to the user based on the recommendation score of the object to be recommended.

[0141] In some example embodiments of the present disclosure, the device 600 also includes: a weight initialization module, which is used to initialize the weights corresponding to each of the business objectives; a weight adjustment module, which is used to adjust the weights corresponding to each of the business objectives in a step-by-step manner, and obtain a set of weights after each adjustment; an adjustment score determination module, which is used to determine the recommendation score corresponding to each set of weights after the weight adjustment; a difference determination module, which is used to determine the difference between the recommendation score and the expected recommendation score, and the expected recommendation score is the expected score of the object to be recommended in the current business scenario; and a weight selection module, which is used to use at least one set of weights corresponding to the recommendation score whose difference is less than a predetermined threshold as the weight of each of the business objectives of the object to be recommended.

[0142] In some example embodiments of the present disclosure, the input layer of the multi-objective recommendation model includes a first network layer and a second network layer, the input features further include cross features, and the input feature extraction module 610 includes: an original feature extraction unit, used to obtain original features from the user information of the user and the object information of the object to be recommended through the first network layer, the original features including the user features and the object features; a cross feature extraction unit, used to obtain the cross features from the user information of the user and the object information of the object to be recommended through the second network layer; a first splicing processing unit, used to splice the original features and the cross features to generate the input features.

[0143] In some example embodiments of the present disclosure, the first network layer is a multi-layer perceptron network, the second network layer is an inner product-based neural network IPNN layer, and the cross feature extraction unit is further used to: obtain user features and object features from the user information of the user and the object information of the object to be recommended through the IPNN layer; perform inner product processing on the user features and / or the object features to generate cross features.

[0144] In some example embodiments of the present disclosure, the hidden layer of the multi-objective recommendation model includes multiple sub-networks and multiple gating layers, and the target score determination module 615 includes: a sub-feature determination module, used to determine the sub-features corresponding to the sub-networks in the input features through each of the sub-networks; a weight determination unit, used to determine the weights of each of the sub-features corresponding to the business goal through the gating layer corresponding to the business goal; a weighted processing unit, used to perform weighted processing on each of the sub-features based on the weights of the sub-features to obtain target features corresponding to the business goals; and a score determination unit, used to determine the target scores of each business goal based on the target features corresponding to each of the business goals.

[0145] In some example embodiments of the present disclosure, the sub-network includes a first cross-feature extraction network and a second cross-feature extraction network, and the sub-feature determination module is further used to: extract high-order cross-features corresponding to the sub-network from the input features through the first cross-feature extraction network; extract second-order cross-features corresponding to the sub-network from the input features through the second cross-feature extraction network, wherein the order of the high-order cross-features is greater than the order of the second-order cross-features; and concatenate the high-order cross-features and the second-order cross-features to generate sub-features corresponding to the sub-network.

[0146] In some example embodiments of the present disclosure, the first cross feature extraction network is a multi-layer perceptron network, and the second cross feature extraction network is a factorization machine network.

[0147] In some example embodiments of the present disclosure, the device 600 also includes: a sample acquisition module, used to acquire sample data, the sample data including sample features and sample labels, the sample features including user features of sample users and object features of sample objects, and the sample labels including actual target scores for each business target of the sample objects; a predicted score determination module, used to input the sample data into the multi-objective recommendation model to determine the predicted target scores for each business target of the sample objects; a loss function determination module, used to determine the loss function corresponding to the business target based on the difference between the actual target score and the predicted target score; and a parameter adjustment module, used to adjust the parameters of the multi-objective recommendation model based on the loss function.

[0148] In some example embodiments of the present disclosure, the recommendation score determination module 620 is further configured to: determine the recommendation score of the object to be recommended by using a Sigmoid activation function based on a result of weighted processing.

[0149] In some example embodiments of the present disclosure, the multi-objective recommendation model is a neural network model based on soft parameter sharing.

[0150] Exemplary Computing Devices

[0151] After introducing the method, medium, and apparatus of the exemplary embodiment of the present disclosure, next, a computing device according to another exemplary embodiment of the present disclosure is introduced.

[0152] Those skilled in the art will appreciate that various aspects of the present disclosure may be implemented as systems, methods or program products. Therefore, various aspects of the present disclosure may be specifically implemented in the following forms, namely: complete hardware implementation, complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to herein as "circuits", "modules" or "systems".

[0153] In some possible embodiments, a computing device according to an embodiment of the present disclosure may include at least one processor and at least one memory. The memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the multi-target recommendation method according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processor may execute the following steps: Figure 2The steps shown in: Step S210, extracting input features from the user information of the user and the object information of the object to be recommended, the input features include: user features and object features; Step S220, based on the input features, determining the target score of each business goal among multiple business goals of the object to be recommended through a multi-objective recommendation model, the multi-objective recommendation model is a neural network model for determining the target score of each business goal of the object to be recommended; Step S230, weighting the target score of each business goal based on the weight corresponding to each business goal, and determining the recommendation score of the object to be recommended based on the result of the weighted processing, the weight is determined according to the expected score of each business goal of the object to be recommended in the current business scenario; recommending the object to be recommended to the user based on the recommendation score of the object to be recommended. For another example, the processor may also execute the following Figure 4 Follow the steps shown in .

[0154] Refer to the following Figure 7 An electronic device 700 according to an exemplary embodiment of the present disclosure is described. Figure 7 The electronic device 700 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0155] like Figure 7 As shown, the electronic device 700 is in the form of a general computing device. The components of the electronic device 700 may include but are not limited to: at least one processing unit 710, at least one storage unit 720, and a bus 730 connecting different system components (including the storage unit 720 and the processing unit 710).

[0156] Bus 730 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures.

[0157] The storage unit 720 may include a readable medium in the form of a volatile memory, such as a RAM (Random Access Memory) 721 and / or a cache memory 722 , and may further include a ROM (Read-Only Memory) 723 .

[0158] The storage unit 720 may also include a program / utility 725 having a set (at least one) of program modules 724, such program modules 724 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0159] The electronic device 700 may also communicate with one or more external devices 740 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 700, and / or communicate with any device that enables the electronic device 700 to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface 750. Furthermore, the electronic device 700 may also communicate with one or more networks (e.g., local area networks, wide area networks, and / or public networks, such as the Internet) via a network adapter 760. As shown, the network adapter 760 communicates with other modules of the electronic device 700 via a bus 730. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID (Redundant Arrays of Independent Disks, independent redundant disk array) systems, tape drives, and data backup storage systems, etc.

[0160] It should be noted that although several units or sub-units of the multi-objective recommendation device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to an embodiment of the present disclosure, the features and functions of two or more modules or units described above may be embodied in one module or unit. Conversely, the features and functions of one module or unit described above may be further divided into multiple modules or units for embodiment.

[0161] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0162] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A multi-objective recommendation method, characterized in that: include: Extracting input features from the user information of the user and the object information of the object to be recommended, the input features including: user features and object features; the input features also include cross features; Based on the input features, a target score for each of the multiple business targets for the object to be recommended is determined through a multi-target recommendation model, and the multi-target recommendation model is a neural network model for determining the target score for each of the multiple business targets for the object to be recommended; the hidden layer of the multi-target recommendation model includes multiple sub-networks and multiple gating layers, and the method of determining the target score for each of the multiple business targets for the object to be recommended through the multi-target recommendation model includes: determining, through each of the sub-networks, a sub-feature in the input features corresponding to the sub-network; determining, through the gating layer corresponding to the business target, the weight of each of the sub-features corresponding to the business target; performing weighted processing on each of the sub-features based on the weight of the sub-feature to obtain a target feature corresponding to the business target; and determining the target score for each business target based on the target feature corresponding to each of the business targets; Performing weighted processing on the target scores of the business targets based on the weights corresponding to the business targets, and determining the recommendation scores of the objects to be recommended based on the results of the weighted processing, wherein the weights are determined according to the expected scores of the business targets of the objects to be recommended in the current business scenario; Initialize the weights corresponding to each of the business objectives; adjust the weights corresponding to each of the business objectives in a step-by-step manner, and obtain a set of weights after each adjustment; determine the recommendation score corresponding to each set of weights after the weight adjustment; determine the difference between the recommendation score and the expected recommendation score, the expected recommendation score being the expected score of the object to be recommended in the current business scenario; use at least one set of weights corresponding to the recommendation score whose difference is less than a predetermined threshold as the weight of each of the business objectives of the object to be recommended; The object to be recommended is recommended to the user based on the recommendation score of the object to be recommended.

2. The method according to claim 1, characterized in that The input layer of the multi-objective recommendation model includes a first network layer and a second network layer, and the step of extracting input features from the user information of the user and the object information of the object to be recommended includes: Acquire original features from the user information of the user and the object information of the object to be recommended through the first network layer, wherein the original features include the user features and the object features; Acquiring the cross-feature from the user information of the user and the object information of the object to be recommended through the second network layer; The original features and the cross features are concatenated to generate the input features.

3. The method according to claim 2, characterized in that The first network layer is a multi-layer perceptron network, the second network layer is an inner product-based neural network IPNN layer, and the cross-features are obtained from the user information of the user and the object information of the object to be recommended through the second network layer, including: Acquire user features and object features from the user information of the user and the object information of the object to be recommended through the IPNN layer; Inner product processing is performed on the user features and / or the object features to generate cross features.

4. The method according to claim 1, characterized in that: The sub-network includes a first cross feature extraction network and a second cross feature extraction network, and determining the sub-features corresponding to the sub-networks in the input features through the sub-networks includes: Extracting high-order cross features corresponding to the sub-network from the input features through the first cross feature extraction network; Extracting, by the second cross feature extraction network, a second-order cross feature corresponding to the sub-network from the input feature, wherein the order of the high-order cross feature is greater than the order of the second-order cross feature; The high-order cross features and the second-order cross features are concatenated to generate sub-features corresponding to the sub-network.

5. The method according to claim 4, characterized in that The first cross feature extraction network is a multi-layer perceptron network, and the second cross feature extraction network is a factorization machine network.

6. The method according to claim 1, characterized in that The method further comprises: Acquire sample data, where the sample data includes sample features and sample labels, where the sample features include user features of sample users and object features of sample objects, and the sample labels include actual target scores for each business target of the sample objects; Inputting the sample data into the multi-objective recommendation model to determine a predicted target score for each business target of the sample object; Determining a loss function corresponding to the business objective based on a difference between the actual target score and the predicted target score; The parameters of the multi-objective recommendation model are adjusted based on the loss function.

7. The method according to any one of claims 1 to 6, characterized in that The determining the recommendation score of the object to be recommended based on the result of weighted processing includes: Based on the result of the weighted processing, the recommendation score of the object to be recommended is determined by a Sigmoid activation function.

8. The method according to claim 7, characterized in that The multi-objective recommendation model is a neural network model based on soft parameter sharing.

9. A multi-objective recommendation device, characterized in that: include: An input feature extraction module is used to extract input features from the user information of the user and the object information of the object to be recommended, wherein the input features include: user features and object features; the input features also include cross features; A target score determination module is used to determine the target score of each business target among the multiple business targets of the object to be recommended through a multi-target recommendation model based on the input features, wherein the multi-target recommendation model is a neural network model for determining the target score of each business target of the object to be recommended; the hidden layer of the multi-target recommendation model includes multiple sub-networks and multiple gating layers, and the target score determination module includes: a sub-feature determination module, which is used to determine the sub-features corresponding to the sub-network in the input features through each of the sub-networks; a weight determination unit, which is used to determine the weights of each of the sub-features corresponding to the business target through the gating layer corresponding to the business target; a weighted processing unit, which is used to perform weighted processing on each of the sub-features based on the weights of the sub-features to obtain the target features corresponding to the business target; a score determination unit, which is used to determine the target score of each business target based on the target features corresponding to each of the business targets; A recommendation score determination module, used to perform weighted processing on the target scores of the respective business targets based on the weights corresponding to the respective business targets, and determine the recommendation score of the object to be recommended based on the result of the weighted processing, wherein the weight is determined according to the expected scores of the respective business targets of the object to be recommended in the current business scenario; A weight initialization module is used to initialize the weights corresponding to each of the business objectives; a weight adjustment module is used to adjust the weights corresponding to each of the business objectives in a step-by-step manner, and obtain a set of weights after each adjustment; an adjustment score determination module is used to determine the recommendation score corresponding to each set of weights after the weight adjustment; a difference determination module is used to determine the difference between the recommendation score and the expected recommendation score, and the expected recommendation score is the expected score of the object to be recommended in the current business scenario; a weight selection module is used to use at least one set of weights corresponding to the recommendation score whose difference is less than a predetermined threshold as the weight of each of the business objectives of the object to be recommended; A recommendation module is used to recommend the object to be recommended to the user based on the recommendation score of the object to be recommended.

10. The device according to claim 9, characterized in that The input layer of the multi-objective recommendation model includes a first network layer and a second network layer, and the input feature extraction module includes: An original feature extraction unit, configured to obtain original features from the user information of the user and the object information of the object to be recommended through the first network layer, wherein the original features include the user features and the object features; A cross-feature extraction unit, configured to obtain the cross-feature from the user information of the user and the object information of the object to be recommended through the second network layer; The first splicing processing unit is used to perform splicing processing on the original features and the cross features to generate the input features.

11. The device according to claim 10, characterized in that The first network layer is a multi-layer perceptron network, the second network layer is an inner product-based neural network IPNN layer, and the cross feature extraction unit is further used for: Acquire user features and object features from the user information of the user and the object information of the object to be recommended through the IPNN layer; Inner product processing is performed on the user features and / or the object features to generate cross features.

12. The device according to claim 9, characterized in that The sub-network includes a first cross feature extraction network and a second cross feature extraction network, and the sub-feature determination module is further used for: Extracting high-order cross features corresponding to the sub-network from the input features through the first cross feature extraction network; Extracting, by the second cross feature extraction network, a second-order cross feature corresponding to the sub-network from the input feature, wherein the order of the high-order cross feature is greater than the order of the second-order cross feature; The high-order cross features and the second-order cross features are concatenated to generate sub-features corresponding to the sub-network.

13. The device according to claim 12, characterized in that The first cross feature extraction network is a multi-layer perceptron network, and the second cross feature extraction network is a factorization machine network.

14. The device according to claim 9, characterized in that The device also includes: A sample acquisition module, used to acquire sample data, wherein the sample data includes sample features and sample labels, wherein the sample features include user features of sample users and object features of sample objects, and the sample labels include actual target scores for each business target of the sample objects; A prediction score determination module, used to input the sample data into the multi-objective recommendation model to determine the prediction target score for each business target of the sample object; A loss function determination module, configured to determine a loss function corresponding to the business objective based on a difference between the actual target score and the predicted target score; A parameter adjustment module is used to adjust the parameters of the multi-objective recommendation model based on the loss function.

15. The device according to any one of claims 9 to 14, characterized in that The recommendation score determination module is also used for: Based on the result of the weighted processing, the recommendation score of the object to be recommended is determined by a Sigmoid activation function.

16. The device according to claim 15, characterized in that The multi-objective recommendation model is a neural network model based on soft parameter sharing.

17. A computing device comprising: A processor and a memory, wherein the memory stores executable instructions, and the processor is used to call the executable instructions stored in the memory to execute the method according to any one of claims 1 to 8.

18. A medium having a program stored thereon, wherein when the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images