Product recommendation based on target recommendation model and model training method and device

Abstract

Embodiments of the present application provide a product recommendation method and device based on a target recommendation model and a model training method and device, the target recommendation model comprising a plurality of sub-models, different sub-models corresponding to different business scenarios; the method comprises: inputting scene features of a target business scenario in which a target user is located and first general features related to a product to be recommended into the target recommendation model; obtaining a first recommendation score output by the target recommendation model based on the scene features and the first general features, the first recommendation score being obtained based on second recommendation scores output by each of the plurality of sub-models, the second recommendation score output by one sub-model being obtained by the sub-model based on the first general features and scene features of a business scenario corresponding to the sub-model; and recommending the product to be recommended to the target user based on the first recommendation score. The present application can be applied to a plurality of different business scenarios without the need to train different recommendation models for different business scenarios.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and more specifically, the embodiments of the present invention relate to a product recommendation and model training method and apparatus based on a target recommendation model. Background Technology

[0002] This section is intended to provide background or context for embodiments of the invention as set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] Product recommendations typically involve using a pre-trained recommendation model to select products from a massive database based on pre-defined business objectives (such as click-through rate or conversion rate) and then recommending these products to users. Summary of the Invention

[0004] Existing technologies require training different recommendation models for different business scenarios, which greatly increases the resource and manpower costs of model operation and maintenance, and has poor scalability. In addition, existing recommendation models are difficult to provide stable recommendation results throughout the entire lifecycle of a business.

[0005] To address at least one of the above problems, this invention provides an improved product recommendation and model training method and apparatus based on a target recommendation model.

[0006] In a first aspect of the present invention, a product recommendation method based on a target recommendation model is provided. The target recommendation model includes multiple sub-models, each corresponding to a different business scenario. The method includes: inputting scenario features of the target business scenario in which the target user is located and a first general feature related to the product to be recommended into the target recommendation model; obtaining a first recommendation score output by the target recommendation model based on the scenario features and the first general feature, wherein the first recommendation score is obtained based on a second recommendation score output by each of the multiple sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general feature and the scenario features of the business scenario corresponding to the sub-model; and recommending the product to be recommended to the target user based on the first recommendation score.

[0007] In one embodiment of the present invention, the first general feature includes at least one of the following: user features of the target user; product features of the product to be recommended; contextual features when recommending the product to the target user; and historical interaction behavior features of the target user with the product to be recommended.

[0008] In another embodiment of the present invention, the scenario features of the target business scenario include at least one of the following: the name of the target business scenario; the scenario business attributes of the target business scenario; and the user category features corresponding to the target business scenario.

[0009] In another embodiment of the present invention, before inputting the scene features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model, the method further includes: inputting the scene features of multiple business scenarios and the second general features related to the first historical recommended product into an original recommendation model, the original recommendation model including multiple original sub-models; obtaining a third recommendation score output by the original recommendation model based on the scene features and the second general features, wherein the third recommendation score is obtained based on a fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scenario corresponding to the original sub-model; training the original recommendation model based on the third recommendation score to obtain the target recommendation model.

[0010] In another embodiment of the present invention, training the original recommendation model based on the third recommendation score includes: determining the probability that the first historical recommended product belongs to a positive sample based on the third recommendation score; training the original recommendation model based on the probability and the true sample category of the first historical recommended product, wherein the true sample category is determined based on business objectives and is used to characterize whether the first historical recommended product is a positive sample.

[0011] In another embodiment of the present invention, each sub-model includes a first feature extraction unit and a second feature extraction unit; each sub-model outputs a second recommendation score based on the following method: the first feature extraction unit extracts features from the first general feature to obtain a first output feature vector; wherein, the first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; the second feature extraction unit extracts features from the scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector; and the first output feature vector and the second output feature vector are crossed element-wise to obtain the second recommendation score output by the sub-model.

[0012] In another embodiment of the present invention, the method further includes: training an original recommendation model comprising multiple original sub-models based on scenario features of multiple business scenarios and second general features related to a first historical recommended product to obtain the target recommendation model; wherein, when the input scenario features include scenario features of a first business scenario, the model parameters of the first feature extraction units in other original sub-models besides the first original sub-model are fixed, and the model parameters of the first feature extraction units in the first original sub-model are adjusted, wherein the first original sub-model is the original sub-model corresponding to the first business scenario; when the scenario features of any business scenario are input, the model parameters of the second feature extraction units in each of the multiple original sub-models are adjusted.

[0013] In another embodiment of the present invention, the first recommendation score is obtained by weighted summation of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario.

[0014] In another embodiment of the present invention, the weight corresponding to a business scenario is obtained based on the real-time traffic distribution of the business scenario. The real-time traffic distribution is obtained by updating the initial traffic distribution of the business scenario using the number of times the recommended products in the business scenario are recommended and the number of times the business goals are achieved within a preset time window. The initial traffic distribution of the business scenario is obtained based on the historical business goals of the business scenario.

[0015] In another embodiment of the present invention, the real-time traffic distribution is a β distribution; the weight corresponding to a business scenario is determined based on the α parameter and β parameter of the real-time traffic distribution of the business scenario; wherein, the β parameter of the real-time traffic distribution of the business scenario is determined based on the β parameter of the initial traffic distribution and the number of times the recommended product in the business scenario is recommended but fails to achieve the business goal; the α parameter of the real-time traffic distribution of the business scenario is determined based on the product of the number of times the recommended product in the business scenario achieves the business goal within a preset time window and the number of positive feedbacks corresponding to the business goal in the business scenario, and the α parameter of the initial traffic distribution.

[0016] In another embodiment of the present invention, the target recommendation model further includes an auxiliary model, and the weights corresponding to each business scenario are obtained by the auxiliary model; the method further includes: training an original recommendation model including multiple original sub-models and an original auxiliary model to obtain the target recommendation model; wherein, during the training of the multiple original sub-models, the weights output by the original auxiliary model are fixed.

[0017] In another embodiment of the present invention, the method further includes: for each of a plurality of second historical recommended products, obtaining a fifth recommendation score of the second historical recommended product output by the target recommendation model based on a third general feature of the second historical recommended product and scene features of a plurality of business scenarios; extracting positive and negative samples from the plurality of second historical recommended products, and testing the target recommendation model based on the fifth recommendation score of the positive samples and the fifth recommendation score of the negative samples; if the test passes, performing the step of inputting the scene features of the target business scenario in which the target user is located and the first general feature related to the product to be recommended into the target recommendation model.

[0018] In a second aspect of the present invention, a product recommendation method based on a target recommendation model is provided, the target recommendation model including a main model and an auxiliary model; the method includes: inputting a first general feature related to the product to be recommended into the target recommendation model; obtaining a first recommendation score output by the target recommendation model based on the first general feature; wherein the first recommendation score is obtained by weighting a second recommendation score output by the main model using weights output by the auxiliary model, the weights being determined based on the real-time traffic distribution of the target business scenario in which the target user is located; and recommending the product to be recommended to the target user based on the first recommendation score.

[0019] In one embodiment of the present invention, the real-time traffic distribution of the target business scenario is determined based on the following method: the initial traffic distribution of the target business scenario is determined based on the historical business objectives of the target business scenario; the initial traffic distribution of the business scenario is updated based on the number of times the recommended products in the business scenario are recommended and the number of times the business objectives are achieved within a preset time window, so as to obtain the real-time traffic distribution of the business scenario.

[0020] In another embodiment of the present invention, the method further includes: training the original recommendation model to obtain the target recommendation model; wherein, during the training of the main model, the weights output by the auxiliary model are fixed.

[0021] In another embodiment of the present invention, the main model includes multiple sub-models, each sub-model corresponding to a different business scenario. Each sub-model is used to output a second recommendation score for the product to be recommended in the business scenario corresponding to the sub-model. The weights output by the auxiliary model include the weights corresponding to each business scenario. The first recommendation score is obtained by weighted summation of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario.

[0022] In another embodiment of the present invention, each sub-model includes a first feature extraction unit and a second feature extraction unit; each sub-model outputs a second recommendation score based on the following method: the first feature extraction unit extracts features from the first general feature to obtain a first output feature vector; wherein, the first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; the second feature extraction unit extracts features from the scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector; and the first output feature vector and the second output feature vector are crossed element-wise to obtain the second recommendation score output by the sub-model.

[0023] In another embodiment of the present invention, the method further includes: for each of a plurality of second historical recommended products, obtaining a fifth recommendation score of the second historical recommended product output by the target recommendation model based on a third general feature of the second historical recommended product and scenario features of multiple business scenarios; extracting positive and negative samples from the plurality of second historical recommended products, and testing the target recommendation model based on the fifth recommendation score of the positive samples and the fifth recommendation score of the negative samples; if the test passes, performing the step of inputting a first general feature related to the product to be recommended into the target recommendation model.

[0024] In a third aspect of the present invention, a model training method is provided for training an original recommendation model, the original recommendation model including multiple original sub-models, different original sub-models corresponding to different business scenarios; the method includes: inputting scenario features of multiple business scenarios and second general features related to a first historical recommended product into the original recommendation model; obtaining a third recommendation score output by the original recommendation model based on the scenario features and the second general features, wherein the third recommendation score is obtained based on a fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scenario features of the business scenario corresponding to the original sub-model; training the original recommendation model based on the third recommendation score to obtain a target recommendation model.

[0025] In a fourth aspect of the present invention, a model training method is provided for training an original recommendation model, the original recommendation model including an original main model and an original auxiliary model; the method includes: inputting a second general feature related to a first historical recommended product into the original recommendation model; obtaining a third recommendation score output by the original recommendation model based on the second general feature; wherein the third recommendation score is obtained by weighting a fourth recommendation score output by the original main model using weights output by the original auxiliary model, the weights being determined based on the real-time traffic distribution of the target business scenario in which the target user is located; and training the original recommendation model based on the third recommendation score to obtain a target recommendation model.

[0026] In a fifth aspect of the present invention, a medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described in any embodiment of the present invention.

[0027] In a sixth aspect of the present invention, a product recommendation apparatus based on a target recommendation model is provided. The target recommendation model includes multiple sub-models, each corresponding to a different business scenario. The apparatus includes: a first input module, configured to input scenario features of the target business scenario in which the target user is located and a first general feature related to the product to be recommended into the target recommendation model; a first acquisition module, configured to acquire a first recommendation score output by the target recommendation model based on the scenario features and the first general feature, wherein the first recommendation score is obtained based on a second recommendation score output by each of the multiple sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general feature and the scenario features of the business scenario corresponding to the sub-model; and a first recommendation module, configured to recommend the product to be recommended to the target user based on the first recommendation score.

[0028] In a seventh aspect of the present invention, a product recommendation apparatus based on a target recommendation model is provided, the target recommendation model including a main model and an auxiliary model; the apparatus includes: a second input module for inputting a first general feature related to the product to be recommended into the target recommendation model; a second acquisition module for acquiring a first recommendation score output by the target recommendation model based on the first general feature; wherein the first recommendation score is obtained by weighting a second recommendation score output by the main model using weights output by the auxiliary model, the weights being determined based on the real-time traffic distribution of the target business scenario in which the target user is located; and a second recommendation module for recommending the product to be recommended to the target user based on the first recommendation score.

[0029] In an eighth aspect of the present invention, a model training apparatus is provided for training an original recommendation model, the original recommendation model including multiple original sub-models, different original sub-models corresponding to different business scenarios; the apparatus includes: a third input module for inputting scenario features of multiple business scenarios and second general features related to a first historical recommended product into the original recommendation model; a third acquisition module for acquiring a third recommendation score output by the original recommendation model based on the scenario features and the second general features, wherein the third recommendation score is obtained based on a fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and scenario features of the business scenario corresponding to the original sub-model; and a first training module for training the original recommendation model based on the third recommendation score to obtain a target recommendation model.

[0030] In a ninth aspect of the present invention, a model training apparatus is provided for training an original recommendation model, the original recommendation model including an original main model and an original auxiliary model; the apparatus includes: a fourth input module for inputting a second general feature related to a first historical recommended product into the original recommendation model; a fourth acquisition module for acquiring a third recommendation score output by the original recommendation model based on the second general feature; wherein the third recommendation score is obtained by weighting the fourth recommendation score output by the original main model using weights output by the original auxiliary model, the weights being determined based on the real-time traffic distribution of the target business scenario in which the target user is located; and a second training module for training the original recommendation model based on the third recommendation score to obtain a target recommendation model.

[0031] In a tenth aspect of the present invention, a computing device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any embodiment of the present invention.

[0032] In an eleventh aspect of the present invention, a target recommendation model is provided, comprising: an input layer for inputting scenario features of a target business scenario in which a target user is located and a first general feature related to a product to be recommended; a cross-topology network layer comprising multiple sub-models, each sub-model corresponding to a different business scenario, each sub-model being used to obtain a second recommendation score of the product to be recommended in the business scenario corresponding to the sub-model based on the first general feature and the scenario features of the business scenario corresponding to the sub-model; and an output layer for outputting a first recommendation score of the product to be recommended based on the second recommendation score of the product to be recommended in each business scenario, the first recommendation score being used to recommend the product to the target user.

[0033] In one embodiment of the present invention, the target recommendation model further includes: an auxiliary model for outputting weights corresponding to each business scenario; and an output layer for weighted summation of the second recommendation scores of the product to be recommended in each business scenario based on the weights corresponding to each business scenario, to obtain the first recommendation score of the product to be recommended.

[0034] In another embodiment of the present invention, the auxiliary model includes: an initialization unit, configured to determine the initial traffic of each business scenario based on the historical business objectives of each business scenario; a calculation unit, configured to calculate the number of times the recommended product is recommended and the number of times the business objective is achieved in each business scenario within a preset time window; and an update unit, configured to update the initial traffic distribution of each business scenario based on the number of recommendations and the number of times the business objective is achieved, to obtain the real-time traffic distribution of each business scenario, and to output the weight corresponding to each business scenario based on the real-time traffic distribution of each business scenario.

[0035] In another embodiment of the present invention, each sub-model includes a first feature extraction unit and a second feature extraction unit; the first feature extraction unit is used to extract features from the first general feature to obtain a first output feature vector; wherein, the first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; the second feature extraction unit is used to extract features from the scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector; the second recommendation score output by the sub-model is obtained by element-wise cross-interaction of the first output feature vector and the second output feature vector.

[0036] According to embodiments of the present invention, a product recommendation and model training method and apparatus based on a target recommendation model employs a target recommendation model comprising multiple sub-models to recommend products. Each sub-model processes the scenario features of its corresponding business scenario and a first general feature related to the product to be recommended, obtaining a second recommendation score for the product in the corresponding scenario. A first recommendation score is then obtained based on the second recommendation scores output by each sub-model, and the product is recommended to the target user based on the first recommendation score. The target recommendation model of the present invention can simultaneously extract the differences and commonalities between different business scenarios, thus being applicable to a variety of different business scenarios. Therefore, it eliminates the need to train different recommendation models for different business scenarios, reducing resource overhead and manpower costs for model maintenance, improving the scalability of the target recommendation model, and providing a better user experience.

[0037] Furthermore, in this embodiment of the invention, the second recommendation score is obtained through the main model in the target recommendation model, and the second recommendation score is weighted by the auxiliary model in the target recommendation model after obtaining weights based on the real-time traffic distribution of the target business scenario. This enables the target recommendation model to take into account the traffic distribution changes at different lifecycles of the business scenario, ensuring the real-time performance and stability of the recommendation results and providing users with a better experience. Attached Figure Description

[0038] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of the invention are illustrated in the drawings by way of example and not limitation, wherein: Figure 1 A schematic diagram illustrating an application scenario according to an embodiment of the present invention is shown. Figure 2 A flowchart illustrating a product recommendation method based on a target recommendation model according to an embodiment of the present invention is shown schematically. Figure 3 A schematic diagram illustrating a method of entering a certain business scenario according to an embodiment of the present invention is shown. Figure 4 A schematic diagram of the structure of a sub-model according to an embodiment of the present invention is shown. Figure 5 A schematic diagram of the structure of an auxiliary model according to an embodiment of the present invention is shown. Figure 6 A flowchart illustrating a product recommendation method based on a target recommendation model according to another embodiment of the present invention is shown schematically; Figure 7 A flowchart illustrating a model training method according to an embodiment of the present invention is shown schematically; Figure 8 A flowchart illustrating a model training method according to another embodiment of the present invention is shown schematically; Figure 9 A schematic diagram of a medium according to an embodiment of the present invention is shown; Figure 10 A block diagram of a product recommendation device based on a target recommendation model according to an embodiment of the present invention is shown schematically. Figure 11 A block diagram of a product recommendation device based on a target recommendation model according to another embodiment of the present invention is shown schematically; Figure 12 A block diagram of a model training apparatus according to an embodiment of the present invention is shown schematically; Figure 13 A block diagram of a model training apparatus according to another embodiment of the present invention is shown schematically; Figure 14 A schematic diagram of the structure of a target recommendation model according to an embodiment of the present invention is shown. Figure 15 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts. Detailed Implementation

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

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

[0041] According to embodiments of the present invention, a control method, medium, device, and computing device for an audio player are proposed.

[0042] In this article, it is important to understand that any number of elements in the accompanying figures is for illustrative purposes and not for limitation, and any naming is for distinction only and has no limiting meaning.

[0043] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments. Invention Overview Product recommendations typically involve using a pre-trained recommendation model to filter products from a massive database and then recommending them to users. Different business scenarios often have different business objectives. For example, in one scenario, the objective is to recommend products with significant discounts and low prices to maximize user conversion rates; while in another scenario, the objective is to recommend as many new products as possible to ensure these new products quickly gain sufficient traffic.

[0045] The inventors have discovered that recommendation models need to recommend products to users based on the business objectives of different business scenarios. Since these objectives vary, different recommendation models often need to be trained separately for each scenario, increasing resource and manpower costs for model maintenance and resulting in poor scalability. Furthermore, product traffic often fluctuates significantly throughout the business lifecycle. For example, in the early stages of a business (low-water stage), the number of users is small, resulting in low product traffic; during the online traffic accumulation period, the number of users gradually increases, and traffic grows from small to large; this stage can last for varying lengths and may experience sudden traffic spikes; in the stable period (high-water stage), the traffic proportions of various products are generally relatively fixed. Existing recommendation models struggle to provide stable recommendation results throughout the entire business lifecycle.

[0046] Therefore, this invention employs a target recommendation model comprising multiple sub-models to recommend products. Each sub-model processes the scenario features of its corresponding business scenario and a first general feature related to the product to be recommended, obtaining a second recommendation score for the product in the corresponding scenario. A first recommendation score is then derived based on the second recommendation scores output by each sub-model, and the product is recommended to the target user based on this first recommendation score. This target recommendation model can simultaneously extract the differences and commonalities between different business scenarios, making it applicable to a variety of different business scenarios. Therefore, it eliminates the need to train different recommendation models for different business scenarios, reducing resource and manpower costs for model maintenance, improving the scalability of the target recommendation model, and providing a better user experience.

[0047] Furthermore, this invention obtains the second recommendation score through the main model in the target recommendation model, and then weights the second recommendation score by obtaining weights based on the real-time traffic distribution of the target business scenario through the auxiliary model in the target recommendation model. This enables the target recommendation model to take into account the traffic distribution changes at different lifecycles of the business scenario, ensuring the real-time performance and stability of the recommendation results.

[0048] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention will be described in detail below.

[0049] Application Scenarios Overview First refer to Figure 1 This is a schematic diagram illustrating an application scenario of the present invention. Users can install a client application (e.g., NetEase Yanxuan, Taobao, JD.com, etc.) on a smart terminal 101 (e.g., a mobile phone, tablet computer, etc.) and send a recommendation request to the server through the client. After receiving the product recommendation request, the server can send the product information of product 103 to the display interface 102 of the client for the user to browse. Users can also perform operations such as viewing, favorites, and adding to cart on the display interface 102 for products of interest. For example, in... Figure 1 In the app, users can click on the image of a product named "Pure Cotton T-shirt" to view detailed information about the product, such as color, size, discount, and user reviews.

[0050] Exemplary methods The following is combined Figure 1 Application scenarios, refer to Figure 2 This document describes a product recommendation method based on a target recommendation model according to exemplary embodiments of the present invention. It should be noted that the above application scenarios are shown only to facilitate understanding of the spirit and principles of the invention, and the embodiments of the invention are not limited in any way. Rather, the embodiments of the invention can be applied to any applicable scenario.

[0051] Example 1: like Figure 2 As shown, the product recommendation method based on the target recommendation model of the present invention includes: Step S201: Input the scenario features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model; Step S202: Obtain the first recommendation score output by the target recommendation model based on the scene features and the first general features, wherein the first recommendation score is obtained based on the second recommendation score output by each of the plurality of sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general features and the scene features of the business scenario corresponding to the sub-model; Step S203: Recommend the product to be recommended to the target user based on the first recommendation score.

[0052] In step S201, the target user is the user who needs product recommendations. For example, if a user sends a product recommendation request to the server through a client, that user can be identified as the target user. The target user can be a single user, meaning that products to be recommended can be recommended to each target user individually. Alternatively, the target user can also include multiple users of the same cluster, meaning that the same product to be recommended can be recommended to various target users of the same cluster. Users can be clustered based on user characteristics. User characteristics include, but are not limited to, at least one of the user's age, gender, occupation, etc. For ease of description, the following explanation uses a single target user as an example.

[0053] The business scenario in which the target user is located can be referred to as the target business scenario. Business scenarios can be diverse, including, for example, promotional business scenarios and traffic exploration business scenarios. In a promotional business scenario, the business objective is to recommend products with significant discounts and lower prices to users to maximize user conversion rates. In a traffic exploration business scenario, the business objective is to recommend more new products to users so that these new products can quickly gain sufficient traffic. Furthermore, some business scenarios aim to maximize the number of times users add products to their carts or favorite them. Besides the business scenarios listed above, the business scenarios in this embodiment of the invention may also include other business scenarios, which will not be listed here.

[0054] Different business scenarios can include different scenario characteristics, which correspond to the business mindset and user profile representation of different business scenarios. A scenario characteristic of a business scenario may include, but is not limited to, at least one of the following: the name of the business scenario, the scenario business attribute of the business scenario, and the user category characteristics corresponding to the business scenario. For example, the name of a business scenario may include "Special Price," "Limited-Time Discount," "Super Deal," "You May Also Like," "Recommended for You," etc. Among these, the scenario business attribute of business scenarios named "Special Price," "Limited-Time Discount," and "Super Deal" is a promotional attribute, while the scenario business attribute of business scenarios named "You May Also Like" and "Recommended for You" is a traffic exploration attribute. Different scenario business attributes correspond to different business objectives. For example, the business objective of the aforementioned promotional attribute business scenario is to maximize user conversion rate, while the business objective of the aforementioned traffic exploration attribute business scenario is to enable new products to obtain sufficient traffic as quickly as possible. User category features corresponding to a business scenario can be used to characterize the categories of users within that scenario. For example, users can be categorized based on their occupation into students, white-collar workers, teachers, doctors, etc.; based on their age, they can be categorized into those under 20, 20-30, 30-40, and over 40; and based on their gender, they can be categorized into males and females. In addition, users can be categorized based on other dimensions, or even multiple dimensions simultaneously; these will not be listed here.

[0055] One or more controls corresponding to business scenarios can be set on the client's display interface 102, such as... Figure 3 As shown, the client's display interface 102 includes a "You May Like" control 1021 and a "Super Deal" control 1022, which are used to enter the traffic exploration scenario and the promotion scenario, respectively. After the target user clicks the "You May Like" control 1021, the server can determine that the target user's target business scenario is the traffic exploration scenario.

[0056] When a target user is in a target business scenario, the scenario features of that scenario can be input into the target recommendation model. Each scenario feature can be a feature vector, and these features can be combined into a feature matrix, which is then input into the target recommendation model. The position of each scenario feature within the feature matrix can be fixed. For example, if the scenario features for a traffic exploration scenario are fixed in the first column of the feature matrix, and those for a promotion scenario are fixed in the third column, then when a target user is in a specific target business scenario, the column vector corresponding to that scenario in the feature matrix is ​​non-zero, while the other column vectors can all be zero. The target recommendation model can then input the scenario features of each business scenario into the corresponding sub-model based on their positions within the feature matrix.

[0057] In addition to scenario features, a first general feature related to the product to be recommended can also be input into the target recommendation model. This first general feature is a feature common to all business scenarios and can be obtained through historical data collected by the client (i.e., offline data, which can be data collected on a T+1 day / hourly basis). The first general feature may include, but is not limited to, at least one of the following: user characteristics of the target user, product characteristics of the product to be recommended, contextual features when recommending the product to the target user, and historical interaction behavior characteristics of the target user regarding the product to be recommended. The target user is the recommendation target of the product to be recommended. The user characteristics of the target user are used to describe the user profile and may include the target user's age, gender, occupation, etc. The product characteristics of the product to be recommended may include the category information, price, discount, and positive review rate of the product to be recommended. The contextual features may include the device model of the target user's client, the target user's geographical location, and the time information when the product to be recommended was recommended to the target user. Historical interaction behavior characteristics may include the target user's sequence of actions on the first historical recommended product, such as clicking, searching, adding to favorites, adding to cart, purchasing, and commenting.

[0058] In step S202, the target recommendation model can input scene features from multiple business scenarios, including the target business scenario, into the sub-model corresponding to each business scenario, and obtain the second recommendation score output by each sub-model. Since the target business scenario in which the target user is located is generally a single business scenario, typically only the scene features input to the sub-model corresponding to the target business scenario are non-zero, while the scene features input to other sub-models are all zero. Thus, although each sub-model can output a second recommendation score, the second recommendation score output by the sub-model corresponding to the target business scenario will dominate, making the final first recommendation score output by the target recommendation model primarily determined by the output of the sub-model corresponding to the target business scenario; that is, the first recommendation score is suitable for the target business scenario.

[0059] In addition, the target recommendation model can also input the first general feature into each sub-model. Assuming there are N sub-models, the input of sub-model 1 includes the scenario features of business scenario 1 and the first general feature, the input of sub-model 2 includes the scenario features of business scenario 2 and the first general feature, and so on, the input of sub-model N includes the scenario features of business scenario N and the first general feature.

[0060] In some embodiments, each sub-model includes a first feature extraction unit and a second feature extraction unit. The first feature extraction unit extracts features from a first general feature to obtain a first output feature vector. The second feature extraction unit extracts features from the scene features of the business scenario corresponding to its sub-model to obtain a second output feature vector. Further, before obtaining the first and second output feature vectors, the scene features and the first general feature can undergo preprocessing such as type conversion, normalization, default value processing, and embedding. Type conversion converts features of different types to the same type, for example, converting integer features to floating-point features. Normalization maps each feature to the same value range. Default value processing fills in missing feature values. Embedding converts the first general feature and scene features into vectors of fixed-size, dense numerical values. Compared to one-hot encoding, embedding optimizes the parameter quantity and computational efficiency issues caused by high-dimensional sparsity.

[0061] Both the first feature extraction unit and the second feature extraction unit may include one or more fully connected DNN (Deep Neural Networks) layers, and the number of fully connected layers in both units is the same. For example, in Figure 4 In the figure, the first feature extraction unit and the second feature extraction unit each include three fully connected layers, and each fully connected layer is shown as a square labeled DNN in the figure.

[0062] In some embodiments, the first feature extraction unit may first fuse the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit, and then extract features from the first general feature. In this way, both the unique information of each business scenario and the common information among multiple business scenarios can be modeled simultaneously. The network parameters may include weight parameters and bias parameters. The parameters and output results of each fully connected layer after parameter fusion are as follows: (1) (2) (3) In formula (1), These are the weight parameters of the fully connected layer of the first feature extraction unit. These are the weight parameters of the fully connected layer in the second feature extraction unit. These are the bias parameters of the fully connected layer of the first feature extraction unit. These are the bias parameters for the fully connected layer of the second feature extraction unit. (This is achieved by...) and By cross-referencing, we can obtain the business scenario. Corresponding weight parameters Through the and By cross-referencing, we can obtain the business scenario. Corresponding bias parameters .by Figure 4 Taking the sub-model shown as an example, the weight parameters of the first, second, and third fully connected layers in the first feature extraction unit of the sub-model can be determined according to formula (1). and bias parameters .

[0063] In formula (2), The input features are those of the fully connected layer of the first feature extraction unit. The output features of the fully connected layer of the first feature extraction unit are used. The weight parameters of each fully connected layer will be determined according to formula (1). and bias parameters Then, the weight parameters of this fully connected layer can be... Bias parameters and input features Substituting into formula (2), the output characteristics of the fully connected layer are obtained. The output characteristics of this fully connected layer It can be used as the input feature of the next fully connected layer to continue to be calculated according to formula (2), and finally the output feature of the last fully connected layer of the first feature extraction unit (i.e. the first output feature vector) is obtained.

[0064] On the other hand, suppose the input features and feature mapping relationship of each fully connected layer in the second feature extraction unit are denoted as follows: and Then it can be done through mapping relationships For input features Mapping is performed to obtain the output features of each fully connected layer in the second feature extraction unit. These output features can be used as the input features of the next fully connected layer in the second feature extraction unit for further calculation, until the output features of the last fully connected layer of the second feature extraction unit (i.e., the second output feature vector) are obtained. In formula (3) The output features of the last fully connected layer of the first feature extraction unit. The output features of the last fully connected layer of the second feature extraction unit. For business scenarios The corresponding number of samples, is the activation function. After fusing (element-by-element cross-linking) the output features of the last fully connected layer of the first and second feature extraction units using formula (3), the second recommendation score output by the sub-model can be obtained.

[0065] The second recommendation scores output by each of the other sub-models in the target recommendation model can be obtained in the same way, which will not be elaborated here. Then, the recommendation scores output by each sub-model can be fused to obtain the first recommendation score. For example, the second recommendation scores output by each sub-model can be weighted and summed to obtain the first recommendation score.

[0066] In step S203, products to be recommended can be recommended to the target user based on a first recommendation score. Specifically, a first recommendation score can be determined for each product to be recommended, and the products to be recommended can be ranked based on their first recommendation scores. Products to be recommended are then recommended to the target user sequentially according to the ranking. The higher the first recommendation score of a product to be recommended, the higher its ranking will be; that is, when recommending products to the target user, products with higher first recommendation scores will be given priority.

[0067] In some embodiments, a first recommendation score can be obtained by weighting and summing the second recommendation scores output by the sub-models corresponding to each business scenario using weights corresponding to each business scenario. The weight corresponding to a business scenario reflects the impact of the online traffic distribution of that business scenario on the recommendation results. A larger weight indicates a greater impact of online traffic distribution on the recommendation results, while the impact of historical data (such as the first general feature and scenario features) is smaller; conversely, a smaller weight indicates a smaller impact of online traffic distribution on the recommendation results, while the impact of historical data is greater. By increasing the weights, the ranking results can be adapted to traffic changes at different lifecycle stages. For example, in the low-water stage of the lifecycle, due to low traffic and scarce historical data, the accuracy of the recommendation results output by the target recommendation model based on historical data is low, while the accuracy of recommendations based on online traffic is relatively high. Therefore, in the low-water stage, increasing the weights can enhance the influence of online traffic on the second recommendation score of the business scenario, thereby improving the accuracy of the recommendation results. When the lifecycle is at a high level, the recommendation results output by the target recommendation model are more accurate. Therefore, by reducing the weight, the influence of online traffic on the recommendation results can be reduced, and the recommendation results can be mainly influenced by historical data.

[0068] In some embodiments, the weights can also be weighted using hyperparameters. Hyperparameters characterize the influence of the weights; the larger the hyperparameter, the greater the influence of the weight, and vice versa. Optionally, the weights can be exponentially weighted using hyperparameters. In this case, the first recommendation score... It can be written as: (4) In formula (4), For business scenarios The second recommendation score output by the corresponding sub-model For business scenarios The corresponding weights Hyperparameters are business scenarios. The corresponding index. (Introduction) and It provides personalized weighting for each business scenario. This directly affects the ranking of the second recommendation score in different business scenarios. Then for To control their influence. The larger, The greater the influence, the less... The smaller the influence.

[0069] In some embodiments, the initial traffic distribution for a business scenario can be determined first based on the historical business objectives of that scenario. The initial traffic distribution can be... β The distribution is a continuous probability distribution in the interval (0,1). β The distribution possesses the following characteristics: 1) It has a varied shape and exhibits good distributional fit; 2) Its value range is between (0,1). Therefore... β Distributions can describe various shapes / events within the 0-1 interval, making them suitable for probabilistic modeling of business activities generated in different business scenarios. Based on the business objectives of various business scenarios over the past N days (N can be 7, 15, 30, etc.), and considering the changes in these objectives over this period, the average business objective for each business scenario can be calculated. μ and variance This establishes the initial traffic distribution for the business scenario. Different business scenarios can correspond to different business objectives; for example, click-through rate (CTR) or conversion rate (CVR) can be used. CTR is the product's click-through rate, which is the actual number of clicks divided by the product's impressions. Conversion rate refers to the percentage of clicks that resulted in a conversion. Mathematically, this means... β mean of the distribution and variance The mathematical description is as follows: (5) in, and They represent β Distribution Parameters and Parameters. Equation (5) can be transformed to obtain equation (6), thus obtaining... β Initial distribution parameter and initial parameter And use it as the initial step for various business scenarios. Parameters and initialization Parameters are used to obtain the parameters for various business scenarios. Parameter set and Parameter set Each item in the two parameter sets mentioned above corresponds to a business scenario.

[0070] (6) After establishing the initial traffic distribution, the initial traffic distribution for that business scenario can be updated using the number of times recommended products were suggested and the number of times business goals were achieved within a preset time window (which can be obtained based on real-time online business data). This will yield the real-time traffic distribution for that business scenario. The preset time window can be set according to actual needs, for example, to 12 hours, 1 day, or 3 days.

[0071] Still with β Taking distribution as an example, the real-time traffic distribution in this business scenario... β The parameters can be based on the initial traffic distribution of this business scenario. β The parameters and the number of times the recommended product was recommended in this business scenario without achieving the business objective were determined; the real-time traffic distribution of this business scenario. α The parameters can be based on the product of the number of times the recommended product achieved the business goal in this business scenario within a preset time window and the number of positive feedbacks corresponding to the business goal in this business scenario, as well as the initial traffic distribution of this business scenario. α Parameters determined.

[0072] Let the positive feedback corresponding to the business objective be denoted as . The subscript indicates the business scenario, and the same applies below. The number of times a product is recommended (also known as exposure) but fails to achieve business goals (e.g., product clicks or conversions) is recorded as follows. The number of times a product achieves its business goals is recorded as follows: Aggregation can be performed according to a preset time window W, repeating the following calculation steps: If a product gains exposure but fails to meet business targets, real-time traffic distribution can be monitored. β Parameters updated to the initial traffic distribution β The sum of the number of times parameters and products were recommended but failed to achieve business goals is denoted as: ; If the product achieves its business goals, the real-time traffic distribution can be... α The parameter is updated to the product of the number of times the product achieved its business goals and the number of positive feedback events corresponding to those goals, multiplied by the initial traffic distribution. α The sum of parameters is denoted as .

[0073] Then, the weight corresponding to this business scenario can be determined based on the real-time traffic distribution of this business scenario. In the case of a real-time traffic distribution of... β In the case of distributed traffic, the weight corresponding to a business scenario can be based on the α parameter of the real-time traffic distribution of that business scenario and... β Parameters are determined. Specifically, this can be based on the α parameter of real-time traffic distribution and... β The new mean is obtained by parameter calculation. μand the new mean μ As the weight for the corresponding business scenario.

[0074] In some embodiments, the target recommendation model may include an auxiliary model for outputting weights corresponding to each business scenario. See also Figure 5 The auxiliary model may include an initialization unit for determining the initial traffic of each business scenario based on the historical business objectives of each business scenario; a calculation unit for calculating the number of times the recommended product is recommended and the number of times the business objective is achieved in each business scenario within a preset time window; and an update unit for updating the initial traffic distribution of each business scenario based on the number of recommendations and the number of times the business objective is achieved, to obtain the real-time traffic distribution of each business scenario, and outputting the weight corresponding to each business scenario based on the real-time traffic distribution of each business scenario.

[0075] The first recommended score for business objectives across multiple business scenarios is as follows: (7) in, It is a business scenario s The second recommendation score output by the corresponding sub-model is detailed in formula (3). It is based on the α parameter and real-time traffic distribution. β The mean obtained from parameter calculation, function It is a business scenario s The fusion function for the corresponding second recommendation score and weight is detailed in formula (4). The final fusion result of the fusion function is the first recommendation score for the business objectives in the multi-business scenario.

[0076] In some embodiments, the target recommendation model can be trained based on scenario features of multiple business scenarios and second general features related to a first historical recommended product. Specifically, the scenario features of multiple business scenarios and the second general features related to the first historical recommended product can be input into an original recommendation model comprising multiple original sub-models. The original sub-models included in the original recommendation model correspond one-to-one with the multiple sub-models included in the target recommendation model. Then, a third recommendation score output by the original recommendation model based on the scenario features and the second general features can be obtained, and the original recommendation model can be trained based on the third recommendation score to obtain the target recommendation model. The third recommendation score is obtained based on a fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scenario features of the business scenario corresponding to the sub-model.

[0077] The second general feature can be obtained based on historical data collected from the client, and its category can be the same as the first general feature. It can also include user features, product features, contextual features, and / or historical interaction behavior features, which will not be elaborated further here. Unlike the process of recommending products (i.e., model deployment), during training, scenario features from multiple business scenarios need to be input into the original recommendation model. During forward propagation, the processing of the original recommendation model is similar to that during model deployment. During backpropagation, the model parameters of the original recommendation model can be adjusted based on the third recommendation score output by the original recommendation model.

[0078] Specifically, the probability that the first historical recommended product belongs to a positive sample can be determined based on the third recommendation score; the original recommendation model is then trained based on the probability and the ground truth category of the first historical recommended product. The ground truth category is determined based on business objectives and is used to characterize whether the first historical recommended product is a positive sample. For example, if the business objective is CTR, the first historical recommended product that has been clicked can be identified as a positive sample, and the first historical recommended product that has been exposed but not clicked can be identified as a negative sample. As another example, if the business objective is CVR, the first historical recommended product that has been converted can be identified as a positive sample, and the first historical recommended product that has been exposed but not converted can be identified as a negative sample. The ground truth category of the first historical recommended product can be determined based on a second general feature of the first historical recommended product. Taking CTR as an example, the second general feature may include the target user's historical interaction behavior features with the first historical recommended product. If the historical interaction behavior features include click behavior, then the first historical recommended product is a positive sample; otherwise, the first historical recommended product is a negative sample. After determining the predicted sample category and the true sample category of the first historical recommended product, a loss function (e.g., cross-entropy loss function) can be established. , means as follows: (8) in, For business scenarios samples The input features (including scene features and second general features). This represents the model parameters of the original recommendation model. For the sample The probability of belonging to a positive sample can be determined based on the third recommendation score output by the original recommendation model. Indicates sample Labels used to characterize samples The true sample category, For the sample size, This represents the cross-entropy loss function.

[0079] When each sub-model includes a first feature extraction unit and a second feature extraction unit, during the training of the original recommendation model, if the input scene features include the scene features of the first business scenario, the model parameters of the first feature extraction units in other original sub-models besides the first original sub-model can be fixed (i.e., the parameters of the first feature extraction units in other original sub-models are frozen), and the model parameters of the first feature extraction units in the first original sub-model can be adjusted. Here, the first original sub-model is the original sub-model corresponding to the first business scenario. When the scene features of any business scenario are input, the model parameters of the second feature extraction units in each of the multiple original sub-models are adjusted (i.e., the parameters of the second feature extraction units in each original sub-model are shared). Through parameter freezing and parameter sharing, the differences between various business scenarios can be strengthened, and the final model parameters corresponding to each business scenario can be obtained.

[0080] When the target recommendation model includes an auxiliary model, during the training of multiple original sub-models in the original recommendation model, the weights output by the original auxiliary model in the original recommendation model can be fixed, that is, the weights are frozen.

[0081] In some embodiments, the trained target recommendation model can also be tested. If the test passes, the product recommendation method of this embodiment can be executed using the target recommendation model. If the test fails, the target recommendation model is retrained.

[0082] Specifically, for each of the multiple second historical recommendation products, the fifth recommendation score of the second historical recommendation product output by the target recommendation model based on the third general feature of the second historical recommendation product and the scenario features of multiple business scenarios can be obtained. Positive and negative samples are extracted from the multiple second historical recommendation products, and the target recommendation model is tested based on the fifth recommendation scores of the positive samples and the fifth recommendation scores of the negative samples.

[0083] In the above embodiment, historical data from the most recent N days can be collected as training and testing data, with historical data from the previous N-1 days used as training data and historical data from day N used as testing data. Therefore, the historical data from day N (including the third general features of each second historical recommendation product pushed to the target user on day N and scenario features from multiple business scenarios) can be input into the target recommendation model, and the fifth recommendation score of each second historical recommendation product output by the target recommendation model can be obtained. The method by which the target recommendation model outputs the fifth recommendation score is similar to the method of outputting the first recommendation score, and will not be repeated here. The fifth recommendation scores of each second historical recommendation product can be sorted, and based on the sorting results, AUC (Area Under Curve, the area enclosed by the ROC curve and the coordinate axis) and / or GAUC (Group AUC, i.e., a weighted average of the AUC calculated based on each target user) metrics can be introduced to evaluate the ranking capability of the target recommendation model. Among them, the AUC metric reflects the probability that "when a random pair of positive and negative samples is drawn, the target recommendation model ranks the positive sample before the negative sample," focusing on the target recommendation model's ranking ability across the overall sample space. The GAUC metric is more granular, reflecting the probability that "for each target user, when a random pair of positive and negative samples is drawn, the target recommendation model ranks the positive sample before the negative sample," focusing more on the target recommendation model's ranking ability for each target user. If, for any randomly drawn pair of positive and negative samples, the target recommendation model can always rank the positive sample before the negative sample (i.e., the fifth recommendation score output by the target recommendation model for the positive sample is higher than the fifth recommendation score output for the negative sample), then it indicates that the target recommendation model has a good ranking ability, meaning the recommendation results are relatively accurate.

[0084] This invention utilizes user historical behavior characteristics, user features, product characteristics, and online traffic distribution across multiple scenarios to recommend personalized products that align with the target user's mindset at different lifecycle stages and in different business scenarios, thereby improving business metrics across multiple e-commerce business scenarios. This invention has the following advantages compared to existing technologies: (1) Existing technologies only focus on specific lifecycles (e.g., the stable period or the early stage of a business), and cannot effectively recommend products throughout the entire lifecycle of a business. Moreover, existing technologies are not applicable to some business scenarios during the online traffic accumulation period (traffic changes, sudden increases or decreases, etc.). This invention adapts to the changes in traffic distribution throughout the entire lifecycle of a business based on dynamic weights. It can feed back changes in online traffic to the target recommendation model in real time, adapt to changes in the lifecycle of online businesses, and ensure the real-time feedback of the recommendation scheme and the stability of the recommendation effect.

[0085] (2) Existing technologies only model around a single scenario, which on the one hand ignores the commonalities of multiple business scenarios, making it difficult to train a target recommendation model in low-traffic scenarios; on the other hand, multiple corresponding recommendation models need to be built for multiple business scenarios, increasing the resource overhead and manpower cost of model operation and maintenance, and the scalability is poor. This invention combines multi-scenario data joint modeling, which can learn the differences between different scenarios, take into account the commonalities between scenarios, reduce resource overhead and operation and maintenance costs, and the multi-scenario data increases the amount of sample data, which is more conducive to training a target recommendation model.

[0086] Example 2: See Figure 6 This invention also provides a product recommendation method based on a target recommendation model, wherein the target recommendation model includes a main model and an auxiliary model; the method includes: Step S601: Input the first general feature related to the product to be recommended into the target recommendation model; Step S602: Obtain the first recommendation score output by the target recommendation model based on the first general feature; wherein, the first recommendation score is obtained by weighting the second recommendation score output by the main model using the weights output by the auxiliary model, and the weights are determined based on the real-time traffic distribution of the target business scenario in which the target user is located; Step S603: Recommend the product to be recommended to the target user based on the first recommendation score.

[0087] In step S601, the first general feature may include, but is not limited to, at least one of the following: user characteristics of the target user, product characteristics of the product to be recommended, contextual features when recommending the product to the target user, and historical interaction behavior characteristics of the target user regarding the product to be recommended. A detailed description of the first general feature is provided in Embodiment 1, and will not be repeated here.

[0088] In step S602, on the one hand, the second recommendation score can be output through the main model; on the other hand, the weights can be output through the auxiliary model. Then, the second recommendation score is weighted based on the weights to obtain the first recommendation score.

[0089] The weights can be determined based on the real-time traffic distribution of the target user's target business scenario. The real-time traffic distribution of the target business scenario can be determined as follows: the initial traffic distribution of the target business scenario is determined based on its historical business objectives; the initial traffic distribution is updated based on the number of times recommended products are featured and the number of times business objectives are achieved within a preset time window, thus obtaining the real-time traffic distribution of the business scenario. The methods for determining the real-time traffic distribution and the methods for determining the weights can be found in Example 1, and will not be repeated here.

[0090] In step S603, the products to be recommended can be recommended to the target user based on the first recommendation score. For example, the products to be recommended can be recommended to the target user in descending order of their first recommendation scores.

[0091] The target recommendation model can be pre-trained. Specifically, the second general feature of historical recommended products can be input into the original recommendation model, which includes the original main model and the original auxiliary model, and the third recommendation score output by the original recommendation model can be obtained. This third recommendation score is obtained by weighting the fourth recommendation score output by the original main model based on the weights output by the original auxiliary model. Based on the third recommendation score, the sample category of historical recommended products can be predicted, which characterizes whether a historical recommended product is a positive sample. The second general feature of historical recommended products can also determine the true sample category of historical recommended products. The true sample category can be pre-labeled based on the second general feature; for example, in the case of a business objective of CTR, historical recommended products that have been exposed but not clicked are negative samples, while those that have been clicked are positive samples. A loss function is established based on the sample categories determined by the original recommendation model and the true sample categories of historical recommended products, thereby training the original recommendation model to obtain the target recommendation model. Furthermore, during the training process, the weights output by the original auxiliary model can be frozen.

[0092] In some embodiments, to enable the target recommendation model to be applied to multiple business scenarios, multiple sub-models can be set in the main model, each sub-model corresponding to a business scenario. Each sub-model is used to output a second recommendation score for the product to be recommended in the business scenario corresponding to the sub-model; the weights output by the auxiliary model include the weights corresponding to each of the multiple business scenarios; the first recommendation score is obtained by weighted summing of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario.

[0093] In embodiments including multiple sub-models, the data input to the target recommendation model includes both a first general feature related to the product to be recommended and scene features of the target business scenario in which the target user is located. Each sub-model may include a first feature extraction unit and a second feature extraction unit. Each sub-model outputs a second recommendation score based on the following method: The first feature extraction unit extracts features from the first general feature to obtain a first output feature vector; wherein the first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; the second feature extraction unit extracts features from the scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector; and the first output feature vector and the second output feature vector are cross-referenced element-wise to obtain the second recommendation score output by the sub-model. Specific details of the method embodiment when the target recommendation model includes multiple sub-models can be found in Embodiment 1, and will not be repeated here.

[0094] In some embodiments, the target recommendation model can also be tested. If the test passes, the product recommendation method in Embodiment 2 can be executed. If the test fails, the target recommendation model is retrained. Specifically, for each of the multiple second historical recommendation products, the fifth recommendation score of the second historical recommendation product output by the target recommendation model based on the third general feature of the second historical recommendation product and the scenario features of multiple business scenarios can be obtained. Positive and negative samples are extracted from the multiple second historical recommendation products, and the target recommendation model is tested based on the fifth recommendation scores of the positive samples and the fifth recommendation scores of the negative samples. The testing process can be referred to Embodiment 1, and will not be repeated here.

[0095] Example 3: See Figure 7 This invention also provides a model training method for training an original recommendation model, wherein the original recommendation model includes multiple original sub-models, and different original sub-models correspond to different business scenarios; the method includes: Step S701: Input the scenario features of multiple business scenarios and the second general features related to the first historical recommended product into the original recommendation model; Step S702: Obtain the third recommendation score output by the original recommendation model based on the scene features and the second general features, wherein the third recommendation score is obtained based on the fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scenario corresponding to the original sub-model; Step S703: Train the original recommendation model based on the third recommendation score to obtain the target recommendation model.

[0096] This method embodiment can be used to train the target recommendation model in Embodiment 1. For specific details of the training process, please refer to the embodiment related to the training process in Embodiment 1, which will not be repeated here.

[0097] Example 4: See Figure 8 This invention also provides a model training method for training an original recommendation model, wherein the original recommendation model includes an original main model and an original auxiliary model; the method includes: Step S801: Input the second general feature related to the first historical recommended product into the original recommendation model; Step S802: Obtain the third recommendation score output by the original recommendation model based on the second general feature; wherein, the third recommendation score is obtained by weighting the fourth recommendation score output by the original main model using the weights output by the original auxiliary model, and the weights are determined based on the real-time traffic distribution of the target business scenario in which the target user is located; Step S803: Train the original recommendation model based on the third recommendation score to obtain the target recommendation model.

[0098] This method embodiment can be used to train the target recommendation model in Embodiment 2. For details of the training process, please refer to the embodiment related to the training process in Embodiment 2, which will not be repeated here.

[0099] Exemplary media After introducing the method of exemplary embodiments of the present invention, the following references are made. Figure 9 An example of the medium of an exemplary embodiment of the present invention will be described.

[0100] In this exemplary embodiment, the above method can be implemented by a program product, such as a portable compact disc read-only memory (CD-ROM) including program code, which can run on a device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this document, the readable medium 90 can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0101] The program product can take the form of any combination of one or more readable media. The readable medium 90 can be a readable signal medium or a readable medium in general. The readable medium 90 can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable media (a non-exhaustive list) include: electrical connections with one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0102] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable medium that can send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0103] The program code contained on the readable medium 90 can be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RE, etc., or any suitable combination thereof.

[0104] Program code for performing the operations of this invention can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0105] Exemplary device Having introduced the medium of exemplary embodiments of the present invention, embodiments of the apparatus of exemplary embodiments of the present invention will now be described.

[0106] refer to Figure 10This invention also provides a product recommendation device based on a target recommendation model, wherein the target recommendation model includes multiple sub-models, and different sub-models correspond to different business scenarios; the device includes: The first input module 1001 is used to input the scenario features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model; The first acquisition module 1002 is used to acquire the first recommendation score output by the target recommendation model based on the scene features and the first general features, wherein the first recommendation score is obtained based on the second recommendation score output by each of the plurality of sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general features and the scene features of the business scene corresponding to the sub-model; The first recommendation module 1003 is used to recommend the product to be recommended to the target user based on the first recommendation score.

[0107] See Figure 11 The present invention also provides a product recommendation device based on a target recommendation model, wherein the target recommendation model includes a main model and an auxiliary model; the device includes: The second input module 1101 is used to input the first general feature related to the product to be recommended into the target recommendation model; The second acquisition module 1102 is used to acquire the first recommendation score output by the target recommendation model based on the first general feature; wherein, the first recommendation score is obtained by weighting the second recommendation score output by the main model using the weights output by the auxiliary model, and the weights are determined based on the real-time traffic distribution of the target business scenario in which the target user is located; The second recommendation module 1103 is used to recommend the product to be recommended to the target user based on the first recommendation score.

[0108] See Figure 12 This invention also provides a model training apparatus for training an original recommendation model, wherein the original recommendation model includes multiple original sub-models, and different original sub-models correspond to different business scenarios; the apparatus includes: The third input module 1201 is used to input the scenario features of multiple business scenarios and the second general features related to the first historical recommended product into the original recommendation model; The third acquisition module 1202 is used to acquire the third recommendation score output by the original recommendation model based on the scene features and the second general features, wherein the third recommendation score is obtained based on the fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scene corresponding to the original sub-model; The first training module 1203 is used to train the original recommendation model based on the third recommendation score to obtain the target recommendation model.

[0109] See Figure 13 This invention also provides a model training apparatus for training an original recommendation model, the original recommendation model including an original main model and an original auxiliary model; the apparatus includes: The fourth input module 1301 is used to input the second general feature related to the first historical recommended product into the original recommendation model; The fourth acquisition module 1302 is used to acquire the third recommendation score output by the original recommendation model based on the second general feature; wherein the third recommendation score is obtained by weighting the fourth recommendation score output by the original main model using the weights output by the original auxiliary model, and the weights are determined based on the real-time traffic distribution of the target business scenario in which the target user is located. The second training module 1303 is used to train the original recommendation model based on the third recommendation score to obtain the target recommendation model.

[0110] Exemplary network structure See Figure 14 This invention also provides a target recommendation model, which includes: Input layer 1401 is used to input the scene features of the target business scenario in which the target user is located and the first general features related to the product to be recommended; The cross-topology network layer 1402 includes multiple sub-models 14021, with different sub-models corresponding to different business scenarios. Each sub-model 14021 is used to obtain the second recommendation score of the product to be recommended in the business scenario corresponding to the sub-model 14021 based on the first general feature and the scenario features of the business scenario corresponding to the sub-model 14021. Output layer 1403 is used to output a first recommendation score of the product to be recommended based on the second recommendation score of the product to be recommended in various business scenarios. The first recommendation score is used to recommend the product to be recommended to the target user.

[0111] Assuming there are N sub-models, for simplicity, only the specific structure of sub-model 1 is shown in the figure. The specific structures of sub-model 2, ..., sub-model N are the same as those of sub-model 1 and are not shown in the figure.

[0112] In some embodiments, the target recommendation model further includes: an auxiliary model 1404, used to output the weights corresponding to each business scenario; and an output layer 1403, used to perform a weighted summation of the second recommendation scores of the product to be recommended in each business scenario based on the weights corresponding to each business scenario, to obtain the first recommendation score of the product to be recommended.

[0113] In some embodiments, the auxiliary model 1404 includes: an initialization unit 14041, configured to determine the initial traffic of each business scenario based on the historical business objectives of each business scenario; a calculation unit 14042, configured to calculate the number of times the recommended product is recommended and the number of times the business objective is achieved in each business scenario within a preset time window; and an update unit 14043, configured to update the initial traffic distribution of each business scenario based on the number of recommendations and the number of times the business objective is achieved, to obtain the real-time traffic distribution of each business scenario, and to output the weight corresponding to each business scenario based on the real-time traffic distribution of each business scenario.

[0114] In some embodiments, each sub-model 14021 includes a first feature extraction unit 14021a and a second feature extraction unit 14021b; the first feature extraction unit 14021a is used to extract features from the first general features to obtain a first output feature vector; wherein, the first feature extraction unit 14021a extracts features from the first general features after fusing the network parameters of the first feature extraction unit 14021a and the network parameters of the second feature extraction unit 14021b; the second feature extraction unit 14021b is used to extract features from the scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector; the second recommendation score output by the sub-model 14021 is obtained by element-wise cross-interaction of the first output feature vector and the second output feature vector.

[0115] In some embodiments, the input layer 1401 is further used to perform preprocessing such as type conversion, normalization, and default value handling on the scene features and the first general features. In other embodiments, the target recommendation model further includes an embedding vector layer 1405, which is used to embed the scene features and the first general features output by the input layer 1401 to obtain vectors corresponding to the scene features and vectors corresponding to the first general features, and then outputs them to the corresponding sub-models.

[0116] The functions performed by each layer in the target recommendation model in this embodiment are detailed in the aforementioned method embodiment, and will not be repeated here.

[0117] Exemplary computing device After introducing the methods, media, and apparatus of exemplary embodiments of the present invention, the following references are made. Figure 15 An embodiment of the computing device according to an exemplary embodiment of the present invention will be described.

[0118] Figure 15 The computing device 1500 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of the present invention.

[0119] like Figure 15 As shown, the computing device 1500 is presented in the form of a general-purpose computing device. The components of the computing device 1500 may include, but are not limited to: at least one processing unit 1501, at least one storage unit 1502, and a bus 1503 connecting different system components (including the processing unit 1501 and the storage unit 1502).

[0120] Bus 1503 includes a data bus, a control bus, and an address bus.

[0121] Storage unit 1502 may include readable media in the form of volatile memory, such as random access memory (RAM) 15021 and / or cache memory 15022, and may further include readable media in the form of non-volatile memory, such as read-only memory (ROM) 15023.

[0122] Storage unit 1502 may also include a program / utility 15025 having a set (at least one) of program modules 15024, such program modules 15024 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0123] The computing device 1500 can also communicate with one or more external devices 1504 (such as a keyboard, pointing device, etc.).

[0124] This communication can be performed via input / output (I / O) interface 1505. Furthermore, the computing device 1500 can also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 1506. Figure 15As shown, network adapter 1506 communicates with other modules of computing device 1500 via bus 1503. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with computing device 1500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0125] It should be noted that although several units / modules or sub-units / modules of the audio player's control device are mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0126] Furthermore, although the operations of the method of the present invention are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0127] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A product recommendation method based on a target recommendation model, wherein the target recommendation model includes multiple sub-models, and different sub-models correspond to different business scenarios; the method includes: The scenario characteristics of the target user's target business scenario and the first general features related to the product to be recommended are input into the target recommendation model; Obtain a first recommendation score output by the target recommendation model based on the scene features and the first general features, wherein the first recommendation score is obtained based on the second recommendation score output by each of the plurality of sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general features and the scene features of the business scenario corresponding to the sub-model; Based on the first recommendation score, the product to be recommended is recommended to the target user; The first recommendation score is obtained by weighted summing of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario. The weight corresponding to a business scenario is obtained based on the real-time traffic distribution of the business scenario. The real-time traffic distribution is obtained by updating the initial traffic distribution of the business scenario using the number of times the recommended products are recommended and the number of times the business goals are achieved within a preset time window. The initial traffic distribution of the business scenario is obtained based on the historical business goals of the business scenario.

2. The method according to claim 1, wherein the first general feature includes at least one of the following: The user characteristics of the target user; The product characteristics of the product to be recommended; Contextual features when recommending the product to the target user; The target user's historical interaction behavior characteristics with the product to be recommended.

3. The method according to claim 1, wherein the scenario characteristics of the target business scenario include at least one of the following: The name of the target business scenario; The scenario business attributes of the target business scenario; The user category characteristics corresponding to the target business scenario.

4. The method according to claim 1, further comprising, before inputting the scenario features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model: The scenario features of multiple business scenarios and the second general features related to the first historical recommended product are input into the original recommendation model, which includes multiple original sub-models; Obtain the third recommendation score output by the original recommendation model based on the scene features and the second general features, wherein the third recommendation score is obtained based on the fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scenario corresponding to the original sub-model; The original recommendation model is trained based on the third recommendation score to obtain the target recommendation model.

5. The method according to claim 4, wherein training the original recommendation model based on the third recommendation score comprises: The probability that the first historical recommended product belongs to the positive sample is determined based on the third recommendation score; The original recommendation model is trained based on the probability and the true sample category of the first historical recommended product. The true sample category is determined based on business objectives and is used to characterize whether the first historical recommended product is a positive sample.

6. The method according to claim 1, wherein each sub-model includes a first feature extraction unit and a second feature extraction unit; each sub-model outputs a second recommendation score based on the following method: The first feature extraction unit extracts features from the first general feature to obtain a first output feature vector; wherein... The first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; The second feature extraction unit extracts scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector. The first output feature vector and the second output feature vector are crossed element-wise to obtain the second recommendation score output by the sub-model.

7. The method according to claim 6, further comprising: The target recommendation model is obtained by training an original recommendation model that includes multiple original sub-models based on the scenario features of multiple business scenarios and the second general features related to the first historical recommendation product. Wherein, when the input scene features include the scene features of the first business scene, the model parameters of the first feature extraction unit in the other original sub-models besides the first original sub-model are fixed, and the model parameters of the first feature extraction unit in the first original sub-model are adjusted. The first original sub-model is the original sub-model corresponding to the first business scene. Given the scene features of any business scenario, the model parameters of the second feature extraction unit in each of the multiple original sub-models are adjusted.

8. The method according to claim 1, wherein the real-time traffic distribution is as follows: β Distribution; the weight corresponding to a business scenario is based on the α parameter of the real-time traffic distribution of that business scenario and β Parameters are determined; among them, Real-time traffic distribution in the aforementioned business scenario β The parameters are based on the initial flow distribution. β The parameters and the number of times the recommended product was recommended but failed to achieve the business objective in the business scenario were determined; The α parameter of the real-time traffic distribution in the business scenario is determined based on the product of the number of times the recommended product in the business scenario achieves the business goal within a preset time window and the number of positive feedbacks corresponding to the business goal in the business scenario, as well as the α parameter of the initial traffic distribution.

9. The method according to claim 1, wherein the target recommendation model further includes an auxiliary model, and the weights corresponding to each business scenario are obtained by the auxiliary model; the method further includes: The target recommendation model is obtained by training an original recommendation model that includes multiple original sub-models and original auxiliary models. During the training of the multiple original sub-models, the weights output by the original auxiliary model are fixed.

10. The method according to claim 1, further comprising: For each of the multiple second historical recommendation products, obtain the fifth recommendation score of the second historical recommendation product output by the target recommendation model based on the third general feature of the second historical recommendation product and the scenario features of multiple business scenarios; Positive and negative samples are extracted from the plurality of second historical recommended products, and the target recommendation model is tested based on the fifth recommendation score of the positive samples and the fifth recommendation score of the negative samples. If the test passes, proceed with the step of inputting the scenario features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model.

11. A product recommendation method based on a target recommendation model, wherein the target recommendation model includes a main model and an auxiliary model; the method includes: Input the first general feature related to the product to be recommended into the target recommendation model; Obtain a first recommendation score output by the target recommendation model based on the first general feature; wherein, the first recommendation score is obtained by weighting the second recommendation score output by the main model using the weights output by the auxiliary model, and the weights are determined based on the following method: determining the initial traffic distribution of the target business scenario based on the historical business objectives of the target business scenario; updating the initial traffic distribution of the business scenario based on the number of times the recommended products in the business scenario are recommended and the number of times the business objectives are achieved within a preset time window to obtain the real-time traffic distribution of the business scenario; Based on the first recommendation score, the product to be recommended is recommended to the target user.

12. The method according to claim 11, further comprising: The original recommendation model is trained to obtain the target recommendation model; During the training of the main model, the weights output by the auxiliary model are fixed.

13. The method according to claim 11, wherein the main model includes multiple sub-models, different sub-models correspond to different business scenarios, each sub-model is used to output a second recommendation score for the product to be recommended in the business scenario corresponding to the sub-model; the weights output by the auxiliary model include the weights corresponding to each business scenario in the multiple business scenarios; the first recommendation score is obtained by weighted summation of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario.

14. The method according to claim 13, wherein each sub-model includes a first feature extraction unit and a second feature extraction unit; each sub-model outputs a second recommendation score based on the following manner: The first feature extraction unit extracts features from the first general feature to obtain a first output feature vector; wherein... The first feature extraction unit extracts features from the first general feature after fusing the network parameters of the first feature extraction unit and the network parameters of the second feature extraction unit; The second feature extraction unit extracts scene features of the business scenario corresponding to the sub-model to obtain a second output feature vector. The first output feature vector and the second output feature vector are crossed element-wise to obtain the second recommendation score output by the sub-model.

15. The method according to claim 11, further comprising: For each of the multiple second historical recommendation products, obtain the fifth recommendation score of the second historical recommendation product output by the target recommendation model based on the third general feature of the second historical recommendation product and the scenario features of multiple business scenarios; Positive and negative samples are extracted from the plurality of second historical recommended products, and the target recommendation model is tested based on the fifth recommendation score of the positive samples and the fifth recommendation score of the negative samples. If the test passes, proceed with the step of inputting the first general feature related to the product to be recommended into the target recommendation model.

16. A model training method for training an original recommendation model, wherein the original recommendation model includes multiple original sub-models, and different original sub-models correspond to different business scenarios; the method includes: The scenario features of multiple business scenarios and the second general features related to the first historical recommended product are input into the original recommendation model; Obtain the third recommendation score output by the original recommendation model based on the scene features and the second general features, wherein the third recommendation score is obtained based on the fourth recommendation score output by each of the multiple original sub-models, and the fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scenario corresponding to the original sub-model; The original recommendation model is trained based on the third recommendation score to obtain the target recommendation model; The third recommendation score is obtained by weighting and summing the fourth recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario. The weight corresponding to a business scenario is obtained based on the real-time traffic distribution of the business scenario. The real-time traffic distribution is obtained by updating the initial traffic distribution of the business scenario using the number of times the recommended products are recommended and the number of times the business goals are achieved within a preset time window. The initial traffic distribution of the business scenario is obtained based on the historical business goals of the business scenario.

17. A model training method for training an original recommendation model, wherein the original recommendation model includes an original main model and an original auxiliary model; the method includes: Input the second general feature related to the first historical recommended product into the original recommendation model; Obtain the third recommendation score output by the original recommendation model based on the second general feature; wherein, the third recommendation score is obtained by weighting the fourth recommendation score output by the original main model using the weights output by the original auxiliary model, and the weights are determined based on the following method: determining the initial traffic distribution of the target business scenario based on the historical business objectives of the target business scenario; The initial traffic distribution of the business scenario is updated based on the number of times the recommended products are recommended and the number of times the business goals are achieved within the preset time window, so as to obtain the real-time traffic distribution of the business scenario. The original recommendation model is trained based on the third recommendation score to obtain the target recommendation model.

18. A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method of any one of claims 1 to 17.

19. A product recommendation device based on a target recommendation model, wherein the target recommendation model includes multiple sub-models, and different sub-models correspond to different business scenarios; the device includes: The first input module is used to input the scenario features of the target business scenario in which the target user is located and the first general features related to the product to be recommended into the target recommendation model; The first acquisition module is used to acquire a first recommendation score output by the target recommendation model based on the scene features and the first general features, wherein the first recommendation score is obtained based on the second recommendation score output by each of the plurality of sub-models, and the second recommendation score output by a sub-model is obtained by the sub-model based on the first general features and the scene features of the business scene corresponding to the sub-model; The first recommendation module is used to recommend the product to be recommended to the target user based on the first recommendation score; The first recommendation score is obtained by weighted summing of the second recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario. The weight corresponding to a business scenario is obtained based on the real-time traffic distribution of the business scenario. The real-time traffic distribution is obtained by updating the initial traffic distribution of the business scenario using the number of times the recommended products are recommended and the number of times the business goals are achieved within a preset time window. The initial traffic distribution of the business scenario is obtained based on the historical business goals of the business scenario.

20. A product recommendation device based on a target recommendation model, wherein the target recommendation model includes a main model and an auxiliary model; the device includes: The second input module is used to input the first general feature related to the product to be recommended into the target recommendation model; The second acquisition module is used to acquire a first recommendation score output by the target recommendation model based on the first general feature; wherein, the first recommendation score is obtained by weighting the second recommendation score output by the main model using the weights output by the auxiliary model, and the weights are determined based on the following method: determining the initial traffic distribution of the target business scenario based on the historical business objectives of the target business scenario; updating the initial traffic distribution of the business scenario based on the number of times the recommended products in the business scenario are recommended and the number of times the business objectives are achieved within a preset time window to obtain the real-time traffic distribution of the business scenario; The second recommendation module is used to recommend the product to be recommended to the target user based on the first recommendation score.

21. A model training apparatus for training an original recommendation model, the original recommendation model comprising multiple original sub-models, different original sub-models corresponding to different business scenarios; the apparatus comprising: The third input module is used to input the scenario features of multiple business scenarios and the second general features related to the first historical recommended product into the original recommendation model; The third acquisition module is used to acquire the third recommendation score output by the original recommendation model based on the scene features and the second general features. The third recommendation score is obtained based on the fourth recommendation score output by each of the multiple original sub-models. The fourth recommendation score output by an original sub-model is obtained by the original sub-model based on the second general features and the scene features of the business scenario corresponding to the original sub-model. The first training module is used to train the original recommendation model based on the third recommendation score to obtain the target recommendation model; The third recommendation score is obtained by weighting and summing the fourth recommendation scores output by the sub-models corresponding to each business scenario using the weights corresponding to each business scenario. The weight corresponding to a business scenario is obtained based on the real-time traffic distribution of the business scenario. The real-time traffic distribution is obtained by updating the initial traffic distribution of the business scenario using the number of times the recommended products are recommended and the number of times the business goals are achieved within a preset time window. The initial traffic distribution of the business scenario is obtained based on the historical business goals of the business scenario.

22. A model training apparatus for training an original recommendation model, the original recommendation model comprising an original main model and an original auxiliary model; the apparatus comprising: The fourth input module is used to input the second general feature related to the first historical recommended product into the original recommendation model; The fourth acquisition module is used to acquire the third recommendation score output by the original recommendation model based on the second general feature; wherein, the third recommendation score is obtained by weighting the fourth recommendation score output by the original main model using the weights output by the original auxiliary model, and the weights are determined based on the following method: determining the initial traffic distribution of the target business scenario based on the historical business objectives of the target business scenario; updating the initial traffic distribution of the business scenario based on the number of times the recommended products in the business scenario are recommended and the number of times the business objectives are achieved within a preset time window to obtain the real-time traffic distribution of the business scenario; The second training module is used to train the original recommendation model based on the third recommendation score to obtain the target recommendation model.

23. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method of any one of claims 1 to 17.

Images