Information recommendation method and model training method, device, storage medium and program product

Abstract

Embodiments of the present application provide an information recommendation method and model training method, device, storage medium and program product. In the embodiments of the present application, for the first commodity information executed by the user for the first positive behavior operation, in the case that the first commodity information does not meet the preset recommendation condition, the interest degree of the user for executing the second positive behavior operation on the plurality of candidate commodity information is predicted by using the target prediction model through the positive and negative behavior sequence information of the user, and the third commodity information capable of jointly executing the second positive behavior operation with the first commodity information is recommended to the user according to the interest degree, so as to meet the preset recommendation condition. Wherein, the positive and negative behavior sequence information more accurately captures the positive interest and negative interest, which helps to guide the user to complete the second positive behavior operation (such as order operation), enhances the satisfaction of the user, and thus reduces the probability of executing the negative behavior operation (such as order cancellation operation).

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to an information recommendation method and model training method, device, storage medium and program product. Background Technology

[0002] In e-commerce, the "spend more, save more" promotion is a common marketing activity. By setting certain conditions and discounts for spending a certain amount, users are encouraged to place orders for more products. That is, when a user selects multiple products to place an order, if the total price or quantity of the multiple products meets the set conditions for spending more, they can enjoy a certain discount.

[0003] If the items a user chooses to order do not meet the minimum purchase requirement for discounts, the e-commerce platform can recommend some add-on items that do meet the minimum purchase requirement to encourage the user to place an order for more items. Users will usually choose to add some add-on items to their order to enjoy the discount.

[0004] However, in practice, there are many cases of order cancellations for bundled items. That is, after enjoying the discount, users will cancel the bundled items, which brings additional workload and costs to the platform and merchants. Summary of the Invention

[0005] This application provides an information recommendation method and model training method, device, storage medium, and program product to recommend product information that better meets users' positive behavioral needs, reduce the probability of users performing negative behavioral operations, and reduce the additional workload that negative behavioral operations bring to the platform and merchants.

[0006] This application provides an information recommendation method, comprising: displaying a target page, the target page including first product information of a target user performing a first positive behavior operation, and the target user generating positive behavior sequence information by performing the first positive behavior operation on the first product information, the first positive behavior operation being associated with a second positive behavior operation; if the first product information does not meet the set recommendation conditions, obtaining negative behavior sequence information generated by the target user performing a negative behavior operation on the second product information within a specified time period; based on the positive behavior sequence information and the negative behavior sequence information, calling a target prediction model to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information; and recommending third product information based on the target user's interest in performing a second positive behavior operation on multiple candidate product information, so that the target user can jointly perform a second positive behavior operation on the first product information and the third product information.

[0007] This application also provides a model training method, comprising: acquiring multiple sample information, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information and their corresponding sample interest and sample success rate of performing a second positive behavior operation; wherein, the positive behavior sequence information is generated by performing a first positive behavior operation on the first product information, and the negative behavior sequence information is generated by performing a negative behavior operation on the second product information; based on the positive behavior sequence information and the negative behavior sequence information, calling an initial prediction model and an initial task constraint model to predict the interest and success rate of the multiple candidate product information, obtaining the predicted interest and prediction success rate corresponding to the multiple candidate product information; constructing a prediction loss function and a constraint loss function based on the predicted interest, sample interest, prediction success rate and sample success rate corresponding to the multiple candidate product information; and jointly training the initial prediction model and the initial task constraint model based on the prediction loss function and the constraint loss function to obtain a target prediction model and a target task constraint model.

[0008] This application embodiment also provides an information prediction model, including: an encoding network, an attention network, and a prediction network; the encoding network is used to encode features of positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, short-term behavior sequence information, and multiple candidate product information to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features; the attention network is used to perform attention calculation on the positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features to generate attention features corresponding to multiple candidate product information; the prediction network is used to predict the interest level of the attention features corresponding to multiple candidate product information to obtain the interest level of multiple candidate product information in performing a second positive behavior operation.

[0009] This application also provides an electronic device, including: a memory and a processor; the memory for storing a computer program; and the processor, coupled to the memory, for executing the computer program to implement the steps in the information recommendation method and model training method provided in this application.

[0010] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to implement the steps in the information recommendation method and model training method provided in this application.

[0011] This application also provides a computer program product, including a computer program / instructions, which, when executed by a processor, cause the processor to implement the steps in the information recommendation method and model training method provided in this application.

[0012] In this embodiment, for the first product information in which a user performs a first positive behavioral operation, if the first product information does not meet the preset recommendation conditions, the system analyzes the user's positive and negative behavioral sequence information and uses a target prediction model to predict the user's interest in performing a second positive behavioral operation on multiple candidate products. Based on this, the system recommends a third product information that can jointly perform the second positive behavioral operation with the first product information, so that the first and third product information jointly meet the preset recommendation conditions. The positive and negative behavioral sequence information can more accurately capture the user's positive and negative interests, thereby achieving more personalized product recommendations, helping to guide the user to complete more second positive behavioral operations, and reducing the probability of the user performing negative behavioral operations.

[0013] Optionally, when the second positive action is an order placement action, the method provided in the embodiments of this application can be used to encourage users to place orders, thereby increasing conversion rates, enhancing user satisfaction, and thus reducing order cancellation rates, thereby reducing the extra workload and costs for the platform and merchants. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0015] Figure 1 A flowchart illustrating an information recommendation method provided for an exemplary embodiment of this application;

[0016] Figure 2a A schematic diagram of the structure of a target prediction model provided for an exemplary embodiment of this application;

[0017] Figure 2b A schematic diagram of another target prediction model provided as an exemplary embodiment of this application;

[0018] Figure 2c A schematic diagram of a model training structure provided for an exemplary embodiment of this application;

[0019] Figure 2d A schematic diagram of a model training structure based on contrastive learning, provided for an exemplary embodiment of this application;

[0020] Figure 3A flowchart illustrating a model training method provided for an exemplary embodiment of this application;

[0021] Figure 4 A schematic diagram of the structure of an information recommendation device provided for an exemplary embodiment of this application;

[0022] Figure 5 This is a schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0025] The various models involved in this application (including but not limited to language models or large models) comply with relevant laws and standards.

[0026] In practical applications, there are frequent order cancellations for bundled items, meaning users cancel their orders after enjoying discounts, causing additional workload and costs for platforms and merchants. In this embodiment, for the first product information where the user performs a first positive action, if the first product information does not meet preset recommendation conditions, the system analyzes the user's positive and negative behavior sequence information and uses a target prediction model to predict the user's interest in performing a second positive action on multiple candidate products. Based on this, it recommends a third product information that can be used in conjunction with the first product information to perform the second positive action, ensuring that the first and third product information jointly meet the preset recommendation conditions. The positive and negative behavior sequence information can more accurately capture the user's positive and negative interests, thereby achieving more personalized product recommendations, helping to guide users to complete more second positive actions, and reducing the probability of users performing negative actions.

[0027] Optionally, when the second positive action is an order placement action, the method provided in the embodiments of this application can be used to encourage users to place orders, thereby increasing conversion rates, enhancing user satisfaction, and thus reducing order cancellation rates, thereby reducing the extra workload and costs for the platform and merchants.

[0028] The following describes in detail a solution provided by an embodiment of this application, with reference to the accompanying drawings.

[0029] Figure 1 This is a flowchart illustrating an information recommendation method provided for an exemplary embodiment of this application. Figure 1 As shown, the method includes:

[0030] 101. Display the target page, which includes the first product information of the target user performing the first positive action, and the target user performing the first positive action on the first product information generates positive action sequence information, and the first positive action is associated with the second positive action.

[0031] 102. If the first product information does not meet the set recommendation conditions, obtain the negative behavior sequence information generated by the target user performing negative behavior operations on the second product information within a specified time period;

[0032] 103. Based on the positive and negative behavior sequence information, call the target prediction model to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information;

[0033] 104. Recommend third product information based on the target user's interest in performing a second positive action on multiple candidate product information, so that the target user can jointly perform a second positive action on the first product information and the third product information.

[0034] In this embodiment, from the perspective of the implementation method of the target page, the target page can be an application page or a web page. For example, the target page can be the application page of a target application, which may include, but is not limited to, e-commerce applications, short video applications, live streaming applications, lifestyle service applications, and social applications. As another example, the target application can be implemented as a webpage of a target platform, which may include, but is not limited to, e-commerce platforms, short video platforms, live streaming platforms, lifestyle service platforms, and social platforms. For example, from the perspective of the functionality of the target page, the target page may include, but is not limited to, shopping cart pages and favorites pages.

[0035] In this embodiment, the target page includes product information, which may include, but is not limited to, daily necessities, food, clothing, home appliances and digital products, and home furnishings. Target users can perform positive actions on one or more product listings on the target page. Positive actions refer to interactive operations that reflect a user's positive interests. For example, positive actions may include, but are not limited to, checking boxes, adding to favorites, adding to cart, liking, commenting, sharing, clicking, viewing, purchasing, placing an order, and making payments.

[0036] For ease of distinction and description, positive actions reflecting a user's interest and consideration stage are referred to as the first positive action; product information executing the first positive action is referred to as the first product information; and positive actions during the user's decision-making and action stages are referred to as the second positive action. In other words, the first positive action is associated with the second positive action, and executing the first positive action is followed by executing the second positive action. For example, the first positive action may include, but is not limited to: checking, adding to favorites, adding to cart, liking, commenting, sharing, clicking, and viewing; the second positive action may include, but is not limited to: purchasing, placing an order, and paying.

[0037] In this embodiment, the target user performs a first positive action operation on the first product information, generating positive action sequence information. The positive action sequence information includes the first positive action operation and its corresponding first product information. Specifically, the positive action sequence information may include one or more first positive action operations; for example, it may include a selection operation on product information in a shopping cart, and the first product information for which the selection operation was performed. Alternatively, the positive action sequence information may also include the first product information corresponding to one or more first positive action operations. The first product information may include, but is not limited to, information such as: name, amount, model, size, sales volume, reviews, shipping location, merchant information, and product details.

[0038] In this embodiment, the first product information can participate in marketing activities, with different marketing activities corresponding to different minimum purchase amounts and discount amounts. For example, the quantity of the first product information can be one or more. If the total price of the first product information reaches the minimum purchase amount of a marketing activity, the total price of the first product information can be reduced by the set discount amount when paying for the first product information. For example, the minimum purchase amount for a minimum purchase activity A1 can be 300, and the set discount amount is 40. If the first product information participates in the minimum purchase activity A1 and the total price of the first product information reaches 300, then the amount paid by the user is: the total price of the first product information minus the discount amount. Another minimum purchase activity A2 can have a minimum purchase amount of 200 and a set discount amount of 20. If the first product information participates in the minimum purchase activity and the total price of the first product information reaches 300, then the amount paid by the user is: the total price of the first product information minus the discount amount. The minimum purchase activity can include same-store minimum purchase discounts and cross-store minimum purchase discounts, etc.

[0039] In this embodiment, the first product information may or may not participate in the product information recommendation activity, depending on the merchant to which the first product information belongs. When the first product information participates in a recommendation activity, it can be identified whether the first product information meets the set recommendation conditions. If the first product information does not meet the set recommendation conditions, a third product information that can jointly execute the second positive action is recommended to the target user. For example, the product information recommendation activity may include, but is not limited to, marketing activities. Marketing activities may include, but are not limited to, same-store discount activities or cross-store discount activities.

[0040] For example, taking a product recommendation activity as a marketing activity and the recommendation conditions as marketing conditions, if the total price of the first product information participating in the marketing activity may not reach the set discount amount, then the first product information can be considered not to meet the set marketing conditions. Alternatively, if the total price of the first product information reaches the discount amount, a discount can be set for the total price of the first product information, and it can be further determined whether the remaining portion of the total price exceeding the discount amount reaches the discount amount. If not, then the first product information can be considered not to meet the set marketing conditions. In this embodiment, if the first product information does not meet the set marketing conditions, a third product information suitable for the target user (e.g., a bundled item) can be recommended, so that the first and third product information together meet the set marketing conditions. For example, the total price of the first and third product information is greater than or equal to the discount amount of the marketing activity. When the set marketing conditions are met, the target user is prompted to perform a second positive action on the first and third product information together, which helps guide the user to perform second positive actions on more product information.

[0041] Optionally, when the second positive action is implemented as an order placement action and the negative action is implemented as an order cancellation action, the technical solution of this embodiment can guide users to place orders for more products. In addition, the recommendation of third product information is based on both the user's order placement and order cancellation actions, prioritizing the recommendation of third product information with a lower user cancellation rate. Thus, when the user places an order for both the first and third product information, the cancellation rate for the third product information is lower, thereby reducing the cancellation rate and minimizing the additional workload and costs for the platform and merchants.

[0042] In this embodiment, the implementation method for recommending third product information includes: obtaining negative behavior sequence information generated by a target user performing negative behavior operations on second product information within a specified time period; based on the positive behavior sequence information and the negative behavior sequence information, calling a target prediction model to predict the target user's interest in performing second positive behavior operations on multiple candidate product information; and recommending third product information based on the target user's interest in performing second positive behavior operations on multiple candidate product information, so that the target user can jointly perform second positive behavior operations on first product information and third product information.

[0043] Negative behavioral actions refer to interactive operations that reflect a user's negative interests. For example, negative behavioral actions may include, but are not limited to, order cancellation, product return, marking items as no longer displayed, and marking items as no longer recommended. For ease of distinction and description, the product information executing a negative behavioral action is referred to as the second product information. The negative behavioral sequence information may include both the negative behavioral action and the second product information executing it. The second product information may include, but is not limited to, information such as name, price, model, size, sales volume, reviews, shipping location, and product details. The specified time period is not limited; for example, the specified time period may include, but is not limited to, one month, two months, three months, or six months.

[0044] The target prediction model can be implemented as any model architecture capable of predicting a target user's interest in performing a second positive behavioral operation on multiple candidate product information. In terms of model scale, the target prediction model in this embodiment can be a model with a large parameter scale, such as various large language models, or a traditional neural network model with a relatively small parameter scale; there is no limitation in this regard. In terms of the model architecture used, the target prediction model can be implemented using, but is not limited to, architectures such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Deep Neural Networks (DNN), Attention Networks, and Residual Networks.

[0045] In this embodiment, based on positive and negative behavior sequence information, a target prediction model is invoked to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information. The description of the second positive behavior operation can be found in the aforementioned embodiments and will not be repeated here. Interest level represents the degree of potential purchase intention shown by the target user towards the product information. Interest level can be represented by probability values; for example, 0% to 100% indicates increasingly higher interest. Multiple candidate product information refers to a subset of products that, after initial screening, are considered likely to meet the specific needs or interests of a user. Multiple candidate product information can be selected from a product database based on the target user's browsing history, purchasing behavior, collection preferences, and multi-dimensional information such as the product's attributes (e.g., category, brand, price range), popularity trends, and seasonal factors.

[0046] In this embodiment, for the first product information in which a user performs a first positive behavioral operation, if the first product information does not meet the preset recommendation conditions, the system analyzes the user's positive and negative behavioral sequence information and uses a target prediction model to predict the user's interest in performing a second positive behavioral operation on multiple candidate products. Based on this, the system recommends a third product information that can jointly perform the second positive behavioral operation with the first product information, so that the first and third product information jointly meet the preset recommendation conditions. The positive and negative behavioral sequence information can more accurately capture the user's positive and negative interests, thereby achieving more personalized product recommendations, helping to guide the user to complete more second positive behavioral operations, and reducing the probability of the user performing negative behavioral operations.

[0047] Optionally, when the second positive action is an order placement action, the method provided in the embodiments of this application can be used to encourage users to place orders, thereby increasing conversion rates, enhancing user satisfaction, and thus reducing order cancellation rates, thereby reducing the extra workload and costs for the platform and merchants.

[0048] In one optional embodiment, the recommendation of a third product based on the target user's interest in performing a second positive action on multiple candidate product information is not limited to the implementation of jointly performing a second positive action on the first and third product information. A specific example is provided below: based on the target user's interest in performing a second positive action on multiple candidate product information, a target product is selected from the multiple candidate product information; the number of target product information can be one or more; the target product information is displayed; in response to the operation of selecting a third product information from the target product information, the third product information is displayed on the target page, allowing the target user to jointly perform a second positive action on the first and third product information. For example, if the target page is implemented as a shopping cart page, when the target user selects the third product information from the target product information, the third product information can be added to the shopping cart and displayed on the shopping cart page.

[0049] The method of displaying the target product information is not limited. For example, the target product information can be displayed on the target page, or other pages can be displayed on the target page, and the target product information can be displayed on those other pages. For example, the display methods of other pages may include, but are not limited to, pop-up boxes, drop-down menus, sidebars, and collapsible boxes.

[0050] Specifically, by accurately recommending target products that match the interests of the target user, the likelihood of the target user performing a second positive action can be significantly increased, thereby improving the overall conversion rate. Personalized recommendations and intelligent order combining help target users quickly find product information that may be of interest, reducing search time and providing a smoother and more convenient shopping experience.

[0051] In an optional embodiment, in addition to acquiring positive and negative behavior sequence information, long-term behavior sequence information in a first time dimension and short-term behavior sequence information in a second time dimension corresponding to the target user can also be acquired, where the first time dimension is longer than the second time dimension. For example, the first time dimension can be half a year, one year, or two years, and the second time dimension can be one week, one month, or three months, etc., without limitation. Both long-term and short-term behavior sequence information can include the time and type of the executed behavior operation, the product information corresponding to the behavior operation, and the merchant information corresponding to the product information, etc., wherein the behavior operation can include positive and negative behavior operations.

[0052] The implementation method of predicting a target user's interest in performing a second positive action on multiple candidate product information by calling a target prediction model based on positive and negative behavior sequence information is not limited. Based on the above, an example is provided, including: predicting a target user's interest in performing a second positive action on multiple candidate product information by calling a target prediction model based on positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, and short-term behavior sequence information, such as... Figure 2a As shown.

[0053] This approach utilizes multi-time-dimensional behavioral analysis, considering not only positive (e.g., purchasing, adding to favorites) and negative (e.g., browsing without purchasing or order cancellation) behaviors of target users, but also their behavioral patterns across multiple time dimensions. Long-term behavioral sequence information reflects stable preferences, while short-term behavioral sequence information may capture current changes in interest or immediate needs. Furthermore, by combining this behavioral sequence information (e.g., positive, negative, long-term, and short-term behavioral sequences) with a target prediction model, the accuracy of interest prediction can be improved by predicting the target user's interest in performing a second positive behavioral action (e.g., purchasing) on ​​multiple candidate products.

[0054] Optionally, in addition to considering long-term and short-term behavioral sequence information, user attribute information of the target user can also be considered. This user attribute information may include user profile information, such as, but not limited to, gender, age, user level, nickname, and location. Based on this, the target prediction model can be invoked to predict the target user's interest in performing a second positive behavioral operation on multiple candidate product information, using user attribute information, positive behavioral sequence information, negative behavioral sequence information, long-term behavioral sequence information, and short-term behavioral sequence information. Figure 2b The illustration shows an example of the target prediction model's input information, which includes user attribute information, positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, and short-term behavior sequence information, but it is not limited to this.

[0055] By comprehensively considering multi-dimensional information about target users (including static user attribute information and dynamic behavioral sequences), the preferences and immediate needs of target users can be captured more accurately. This enables the target prediction model to provide more personalized product recommendations that are more in line with the actual interests of target users, thereby significantly improving the accuracy of interest prediction.

[0056] Optionally, the implementation method of calling the target prediction model to predict the target user's interest in performing a second positive action on multiple candidate product information based on positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, and short-term behavior sequence information is not limited.

[0057] The following section uses a target prediction model, comprising a target encoding network, a target attention network, and a target prediction network, as an example. The target encoding network transforms the input information into a form that the target prediction model can process, providing rich feature representations for subsequent attention and prediction networks. The implementation of the target encoding network is not limited. For example, it can include, but is not limited to, embedding layers, convolutional layers, self-attention layers, fully connected layers, and recurrent neural network layers.

[0058] The target attention network utilizes an attention mechanism to identify the portion of the input data relevant to the predicted interest level. This means it can dynamically adjust the evaluation weights of interest in different product information based on information such as the user's short-term behavior, long-term interests, positive behavior sequence information, and negative behavior sequence information. The attention mechanism allows the target prediction model to focus more on information more critical to the current task of predicting interest level, improving the model's performance in predicting interest level and enhancing its interpretability.

[0059] The target prediction network is used to perform the final prediction based on the information extracted and processed by the previous layers. Specifically, it predicts the target user's interest in performing a second positive action on multiple candidate items. The target prediction network determines the quality of the interest output by the model. By integrating information from the target encoding network and the target attention network, the target prediction network can generate more accurate predicted interest levels, thereby improving the performance of the target prediction model.

[0060] Among them, such as Figure 2bAs shown, according to the target encoding network, positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, short-term behavior sequence information, and multiple candidate product information are feature-encoded to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features. According to the target attention network, attention is calculated on the positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features to generate attention features corresponding to multiple candidate product information. According to the target prediction network of the target prediction model, interest prediction is performed on the attention features corresponding to multiple candidate product information to obtain the interest level of multiple candidate product information in performing the second positive behavior operation. Furthermore, the target encoding network can also perform feature encoding on user attribute information to obtain user attribute features.

[0061] Optionally, in the target attention network, a multi-head attention mechanism can be used to generate multiple attention features corresponding to multiple candidate product features, positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, and short-term behavior sequence features. The multi-head attention mechanism allows simultaneous attention to different parts of the input sequence in different representation subspaces, helping to capture complex dependencies within the input sequence and thus capturing richer information, enhancing the expressive power and flexibility of the target prediction model.

[0062] Specifically, for any candidate product feature, the candidate product feature, positive behavior sequence feature, negative behavior sequence feature, long-term behavior sequence feature, and short-term behavior sequence feature are input into the target attention network. The candidate product feature is then subjected to multi-head attention calculations with the positive and negative behavior sequence features, respectively, to obtain the first and second attention features corresponding to the candidate product information. Based on the multi-head attention results between the candidate product feature and the long-term behavior sequence feature, and between the candidate product feature and the short-term behavior sequence feature, a third attention feature corresponding to the candidate product information is generated. Figure 2b The text provides an example demonstration of the first attention feature, the second attention feature, and the third attention feature.

[0063] Optionally, before performing multi-head attention calculation, the method further includes performing self-attention calculation on candidate product features, positive behavior sequence features, and negative behavior sequence features respectively; performing multi-head attention calculation on the self-attention results of candidate product features and the self-attention results of positive behavior sequence features to obtain a first attention feature; and performing multi-head attention calculation on the self-attention results of candidate product features and the self-attention results of negative behavior sequence features to obtain a second attention feature.

[0064] By inputting candidate product features, along with positive, negative, long-term, and short-term behavioral sequence features, into the target attention network for multi-head attention computation, the model can gain a deeper understanding and capture the complex relationship between user interests and product features. This method considers not only positive user preferences (positive behavioral sequence information) but also negative feedback (negative behavioral sequence information), thus providing a more comprehensive perspective for evaluating the target user's interest in different product information.

[0065] The first and second attention features reflect the correlation between candidate items and users' positive and negative behaviors, respectively, helping to identify item features that strongly attract or clearly do not attract the target user's attention. The third attention feature combines long-term and short-term behavioral patterns, enabling the model to simultaneously consider users' stable preferences and their latest changes in interest, improving the adaptability and flexibility of the target prediction model.

[0066] Further optionally, the implementation method for generating the third attention feature corresponding to the candidate product information based on the multi-head attention result between the candidate product features and the long-term behavior sequence features, and the multi-head attention result between the candidate product features and the short-term behavior sequence features, is not limited. An example is provided below: at least a portion of the features that match the candidate product features are selected from the long-term behavior sequence features as updated long-term behavior sequence features; self-attention calculation is performed on the short-term behavior sequence features to obtain the self-attention result; the third attention feature corresponding to the candidate product information is generated based on the multi-head attention result between the updated long-term behavior sequence features and the candidate product features, and the multi-head attention result between the candidate product features and the self-attention result. For example, the third attention feature is obtained by adding the multi-head attention result between the updated long-term behavior sequence features and the candidate product features to the multi-head attention result between the candidate product features and the self-attention result.

[0067] The implementation method of selecting at least a portion of the features that match the category of the candidate product features from the long-term behavioral sequence features is not limited. For example, a filtering module can be used to select at least a portion of the features that match the category of the candidate product features from the long-term behavioral sequence features as the updated long-term behavioral sequence features, such as... Figure 2b As shown. For example, the filtering module can be implemented as a hard search module. Long-term behavioral sequence features and candidate product features can be input into the hard search module for feature matching, obtaining at least some features that match the category of the candidate product features. These at least some features are then used as the updated long-term behavioral sequence features (also known as user activation interest sequences). Hard search is a precise matching method performed in a specific data structure or algorithm, focusing on filtering results that match given conditions (such as candidate product features).

[0068] By selecting long-term behavioral sequence features that match the candidate product features and performing multi-head attention calculations on them, we can more accurately capture users' stable preferences for specific products, which helps to identify users' deep interests and thus improve the accuracy and relevance of the target prediction model.

[0069] Further, optionally, the implementation method of predicting the interest level of attention features corresponding to multiple candidate product information based on the target prediction network of the target prediction model to obtain the interest level of multiple candidate product information to perform the second positive behavior operation is not limited. A specific implementation method is provided below: for any candidate product feature, a concatenated feature corresponding to the candidate product feature is generated based on the first attention feature, the second attention feature, and the third attention feature; feature prediction is performed on the concatenated feature to obtain the interest level corresponding to the candidate product feature.

[0070] The internal structure of the target prediction network is not limited. For example, the target prediction network may include: a target stitching module, a target feature transfer module, a target feature mapping module, and a target nonlinear activation module.

[0071] The implementation method for generating the splicing features corresponding to the candidate product features based on the first attention feature, the second attention feature, and the third attention feature is not limited. For example, the first attention feature, the second attention feature, the third attention feature, the user attribute feature, and the candidate product feature can be input into the target splicing module for splicing to obtain the splicing features corresponding to the candidate product feature.

[0072] The implementation method for predicting the interest level corresponding to the candidate product features by performing feature prediction on the spliced ​​features is not limited. It can be achieved by the cooperation of a target feature transfer module, a target feature mapping module, and a target nonlinear activation module. Specifically, the spliced ​​features are input to the target feature transfer module for feature transfer processing to obtain transferred features; the transferred features are input to the target feature mapping module for feature mapping to obtain mapped features; and the mapped features are input to the target nonlinear activation module for nonlinear transformation to obtain the interest level corresponding to the candidate product features.

[0073] The implementation method of the target feature transfer module is not limited. Any module that can achieve feature transfer is applicable to the embodiments of this application. For example, the target feature transfer module may include, but is not limited to: an adaptive transfer module, a module with transfer function implemented based on CNN, a module with transfer function implemented based on Support Vector Machine (SVM) or Random Forest, etc. Among them, the adaptive transfer module is used to transfer information between multiple sequences, adaptively learning what information needs to be transferred and how much information needs to be transferred at different stages, thereby effectively capturing the correlation between sequences and improving the overall performance of the target prediction model.

[0074] The implementation method of the target feature mapping module is not limited. Any implementation method that can achieve feature mapping is applicable to the embodiments of this application. For example, the target feature mapping module can be implemented by a multilayer perceptron (MLP).

[0075] The target nonlinear activation module can be implemented using various nonlinear activation functions. For example, the nonlinear activation functions that the target nonlinear activation module can use include, but are not limited to, the S-function (sigmoid), the hyperbolic tangent function (Tanh), and the rectified linear unit (ReLU) function.

[0076] Among these features, concatenated features allow the target prediction model to comprehensively consider information from multiple different sources, helping to capture complex user-product interaction patterns and improve the model's performance. By combining concatenated features with the feature transfer and mapping capabilities of the target prediction layer, highly personalized product recommendations can be provided for each user. This not only improves the relevance of recommendations but also enhances the user experience, promoting higher user engagement and satisfaction.

[0077] In an optional embodiment, this application also provides a training method for a target prediction model. This method utilizes a task-constrained model to assist in training the target prediction model. The success rate (e.g., conversion rate or order rate) of performing the second positive action on candidate product information is used as a constraint task. During model training, the model parameters of the initial prediction model are optimized so that the initial prediction model can simultaneously consider the success rate corresponding to the candidate product information and predict the interest level of performing the second positive action on the candidate product information. In this way, candidate product information with higher interest and success rates can be preferentially recommended to the target user, improving the target user's shopping experience and reducing the order cancellation rate.

[0078] Specifically, multiple sample information is acquired, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information, sample interest levels corresponding to the multiple candidate product information, and sample success rates for executing the second positive behavior operation on the multiple candidate product information. Based on the positive and negative behavior sequence information, the initial prediction model and the initial task constraint model are invoked to predict the interest and success rates of the multiple candidate product information, obtaining the predicted interest and prediction success rates corresponding to the multiple candidate product information. Based on the predicted interest, sample interest, prediction success rate, and sample success rate corresponding to the multiple candidate product information, a prediction loss function and a constraint loss function are constructed. Based on the prediction loss function and the constraint loss function, the initial prediction model and the initial task constraint model are jointly trained to obtain the target prediction model and the target task constraint model.

[0079] Among them, the positive behavior sequence information of the sample is generated by the sample user performing the first positive behavior operation on the first product information, and the negative behavior sequence information of the sample is generated by the sample user performing the negative behavior operation on the second product information within a specified time period.

[0080] Here, sample interest can be understood as the actual interest level in the candidate product information. For example, the actual interest level can be determined by the ratio of the number of times a first positive action is performed on the candidate product information to the number of times the candidate product information is displayed (or visited). For instance, one way to implement sample interest is through the actual click-through rate (CTR) of the candidate product information. For example, the actual CTR can be determined by the ratio of the number of clicks (or selections) on the candidate product information to the number of times it is displayed. Sample success rate can be understood as the ratio of the number of times a second positive access action is performed on the candidate product information to the number of times the candidate product information is displayed (or visited). For instance, one way to implement sample success rate is through the actual transaction rate of the candidate product information. For example, the actual transaction rate can be determined by the ratio of the number of actual purchases of the candidate product information to the number of visits to the product page.

[0081] The prediction loss function measures the difference between the predicted interest level and the sample interest level. Commonly used loss functions include mean squared error (MSE) and cross-entropy loss. For example, the prediction loss function can be implemented as a click-through rate (CTR) loss function. The constraint loss function measures the difference between the prediction success rate and the sample success rate. This helps ensure that the target prediction model can not only accurately predict the user's interest level, but also effectively estimate the probability that the task-constrained model will actually perform a second positive action.

[0082] Furthermore, the constraint loss function can be calibrated (Calibrator). Based on the calibrated constraint loss function, the model parameters of the initial inference model can be optimized, which can improve the accuracy of interest prediction. Figure 2c As shown. Calibration is used to adjust or calibrate the constraint loss function to ensure that its prediction results (such as transaction rates) are more accurate, reliable, and meet specific requirements or standards.

[0083] Gradient descent or other optimization algorithms can be used to minimize the sum of the two loss functions mentioned above, thereby simultaneously adjusting the parameters of the initial prediction model and the initial task constraint model. This joint training method allows the two models to complement each other, improving overall performance.

[0084] Optionally, the implementation method of calling the initial prediction model and the initial task constraint model based on the positive and negative behavioral sequence information of the samples to predict the interest and success rate of multiple candidate product information is not limited. The following provides a specific implementation method: based on the positive and negative behavioral sequence information of the samples, the initial prediction model is called to predict the interest of multiple candidate product information; wherein, during the interest prediction process, the initial prediction model can also generate multiple sample migration features corresponding to the multiple candidate product information; based on the multiple sample migration features, the positive and negative behavioral sequence information of the samples, the initial task constraint model is called to predict the success rate of the multiple candidate product information, obtaining the predicted success rate of each candidate product information. By using the sample migration features generated by the initial prediction model, the initial task constraint model can obtain additional information dimensions, thereby making a more accurate judgment when predicting the success rate.

[0085] Among them, such as Figure 2c As shown, the initial prediction model includes a first initial encoding network, an initial attention network, and a first initial prediction network. For details on the functions implemented by the first initial encoding network, initial attention network, and first initial prediction network, please refer to the descriptions of the aforementioned target encoding network, target attention network, and target prediction network. It should be noted that the first initial encoding network, initial attention network, and first initial prediction network have the same model structure as the aforementioned target encoding network, target attention network, and target prediction network, respectively; the difference lies in the parameters of each layer. The first initial prediction network includes: a first concatenation module, a first feature transfer module, a first feature mapping module, and a first nonlinear activation module. Detailed descriptions can be found in the descriptions of the target concatenation module, target feature transfer module, target feature mapping module, and target nonlinear activation module, and will not be repeated here.

[0086] Among them, such as Figure 2c As shown, the initial task constraint model includes: a second initial encoding network and a second initial prediction network. For a detailed introduction to the second initial encoding network, please refer to the first encoding network. The network parameters of the first and second initial encoding networks can be the same or different; this is not limited. Figure 2cThe illustration uses the example of the first and second initial encoding networks being the same encoding network, but it is not limited to this. The second initial prediction network includes: a second concatenation module, a second feature transfer module, a second feature mapping module, and a second nonlinear activation module. For detailed descriptions, please refer to the previous descriptions of the target concatenation module, target feature transfer module, target feature mapping module, and target nonlinear activation module; these will not be repeated here. It should be noted that the second initial encoding network and the second initial prediction network have the same model structure as the aforementioned target encoding network and target prediction network, respectively; the difference lies in the parameters of each layer.

[0087] The first initial encoding network is used to encode the positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, short-term behavior sequence information, and multiple candidate product information of the samples to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features.

[0088] The initial attention network is used to perform attention calculations on the positive and negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features of the samples, so as to generate sample attention features corresponding to multiple candidate product information. The sample attention features corresponding to any candidate product information include the first sample attention feature, the second sample attention feature, and the third sample attention feature.

[0089] The first initial prediction network is used to predict the interest degree of the sample attention features corresponding to multiple candidate product information, so as to obtain the interest degree of the multiple candidate product information to perform the second positive behavior operation.

[0090] Specifically, for any sample candidate product feature, the first sample attention feature, the second sample attention feature, the third sample attention feature, and the sample candidate product feature are input into the first concatenation module for concatenation to obtain the first concatenation feature corresponding to the candidate product feature; the first concatenation feature is input into the first feature transfer module for feature transfer to obtain the first transfer feature corresponding to the sample candidate product feature; the first transfer feature is input into the first feature mapping module for feature mapping to obtain the first mapping feature; the first mapping feature is input into the first nonlinear activation module for nonlinear transformation to obtain the predicted interest degree for performing the second positive behavior operation on the sample candidate product feature.

[0091] Specifically, the first transfer feature corresponding to the sample candidate product features can be provided to the second feature transfer module in the initial task constraint model through the information interaction (info) module.

[0092] In the initial task constraint module, the positive and negative behavior sequence information of the samples are input into the second initial encoding network for feature encoding, resulting in positive and negative behavior sequence features. These features, along with the first transfer features corresponding to the candidate product features, are then input into the second initial prediction network for feature prediction, yielding the prediction success rate for performing the second positive behavior operation on the candidate product features.

[0093] In the second initial prediction network, the positive and negative behavioral sequence features of the samples can be input into the second concatenation module for feature concatenation to obtain the second concatenated features; the second concatenation features and the first transfer features can be input into the second feature transfer module for feature transfer to obtain the second transfer features; the second transfer features can be input into the second feature mapping module for feature mapping to obtain the second mapped features; the second mapped features can be input into the second nonlinear activation module for nonlinear transformation to obtain the prediction success rate corresponding to the sample candidate product features.

[0094] Optionally, for any sample candidate product information, the marketing information related to the marketing activities of the sample candidate product information can also be input into the second initial encoding network for feature encoding to obtain the marketing features corresponding to the sample candidate product information. Here, marketing information refers to price features related to the marketing activities, such as discount amounts or promotional amounts. The sample positive behavior sequence features, sample negative behavior sequence features, marketing features, and the first transfer features corresponding to the sample candidate product features are input into the second initial prediction network for feature prediction to obtain the prediction success rate of performing the second positive behavior operation on the sample candidate product features. Further, in the second initial prediction network, the sample positive behavior sequence features, sample negative behavior sequence features, and marketing features can be input into the second concatenation module for feature concatenation to obtain the second concatenated features.

[0095] In one optional embodiment, compared to the temporal dimension of behavioral sequence information, positive and negative behavioral sequence information are shorter in size. During model training, directly inputting these shorter sequences into the initial prediction and task constraint models may mask their effectiveness, preventing them from fully utilizing their potential. Therefore, contrastive learning techniques can be employed to enhance these shorter sequences, allowing them to play a more significant role in the training of the initial prediction and task constraint models. Contrastive learning is a self-supervised learning method that aims to construct effective feature representations by learning the similarities and differences between positive and negative behavioral sequence information, thus helping to capture high-quality feature representations. Furthermore, contrastive learning involves gradient backpropagation; the positive and negative behavioral sequence features obtained through contrastive learning are more robust and more helpful in predicting interest and success rates.

[0096] Based on this, the model parameters of the first and second initial prediction networks can be optimized using contrastive learning techniques. The implementation methods for optimizing the model parameters of the first and second initial prediction networks are the same or similar. The following description uses the optimization of the model parameters of the first initial prediction network as an example. Specifically, for any sample candidate product feature, an alignment loss function is constructed based on the sample candidate product features, sample positive behavior sequence features, and sample negative behavior sequence features output by the first initial prediction network; data augmentation is performed on the sample positive behavior sequence features and sample negative behavior sequence features respectively to obtain pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups; a contrastive loss function is constructed based on the pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups; the first initial encoding network is iteratively trained based on the alignment loss function and the contrastive loss function to obtain the target encoding network.

[0097] The alignment loss function measures the similarity between the features of candidate items and the features of positive and negative behavioral sequences, respectively. Multiple candidate item information includes positive and negative candidate item information. Positive candidate item information refers to the candidate items for which the user performs positive actions, and negative candidate item information refers to the candidate items for which the user performs negative actions. The alignment loss function is constructed with the objectives of reducing the distance between positive and negative candidate item information and the features of positive behavioral sequences, increasing the distance between positive and negative candidate item information and the features of negative behavioral sequences, and increasing the distance between negative candidate item information and the features of positive behavioral sequences.

[0098] Data augmentation is an important way to improve the generalization ability of a model and increase the diversity of training data. The specific implementation of data augmentation is not limited. For example, implementations include, but are not limited to, random masking, feature-wise masking, and dimension-wise masking. Random masking is used to simulate data loss by randomly selecting elements in a sequence and masking their values ​​(e.g., setting them to 0 or other specific values), helping the model learn more robust feature representations. Feature masking focuses on operating on specific features, selectively masking certain features to enhance the model's dependence on these features or its ability to independently handle missing data. It helps understand which features are most important for model prediction and also enhances the model's performance when faced with incomplete information. Dimension-wise masking is similar to feature masking, but it operates on a specific dimension of the data. For example, in multidimensional time series data, data for a certain time period can be masked, allowing the model to learn how to infer the masked portion from the remaining information.

[0099] During model training, for any batch, one of the above-mentioned data augmentation methods can be randomly selected. Figure 2d The text illustrates three data augmentation methods, using random masking, feature masking, and dimensionality masking as examples. It demonstrates how to perform data augmentation using any one of these methods to hide some features (i.e., mask features) of a sequence (such as positive and negative behavioral sequence features), and then replace these mask features with new features to generate new behavioral sequence features. The original feature is E∈R. L×DThe example is illustrated below, but it is not limited to this. The original features include negative behavioral sequence features and positive behavioral sequence features of the samples. The pseudo-sequence after data augmentation is denoted by E^, and R represents the dimension of the original features, which is L multiplied by D. L can be the training length, and D can be the number of features.

[0100] For example, data augmentation is performed on the positive behavior sequence features of the samples to obtain a pseudo-positive behavior sequence feature group, which includes at least two pseudo-positive behavior sequence features; data augmentation is performed on the negative behavior sequence features of the samples to obtain a pseudo-negative behavior sequence feature group, which includes at least two pseudo-negative behavior sequence features.

[0101] For example, in Figure 2d In this diagram, the sample candidate product features are represented as Ft, the sample positive behavior sequence features are represented as Fp, the sample negative behavior sequence features are represented as Fn, and the pseudo-positive behavior sequence feature group includes pseudo-positive behavior sequence Fp'1 and pseudo-positive behavior sequence Fp'2, and the pseudo-negative behavior sequence feature group includes pseudo-positive behavior sequence Fn'1 and pseudo-positive behavior sequence Fn'2, as an example for illustration.

[0102] Optionally, the implementation method for constructing the contrastive loss function based on the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group is not limited. A specific example is provided below, aiming to reduce the intra-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group, and to increase the inter-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group, to construct the contrastive loss function.

[0103] For example, a contrastive loss function is constructed with the goal of reducing the distance between Fp'1 and Fp'2, and between Fn'1 and Fn'2, and increasing the distance between Fp'1 and Fn'1, Fp'1 and Fn'2, Fp'2 and Fn'1, and Fp'2 and Fn'2. Since the sample positive behavior sequence feature Fp and the sample negative behavior sequence feature Fn are obtained by encoding (or vectorizing) the sample positive behavior sequence information and sample negative behavior sequence information by the first initial encoding network, it is necessary to continuously adjust the model parameters of the first initial encoding network so that the first initial encoding network can generate Fp and Fn that satisfy the alignment loss function and the contrastive loss function, making the two features Fp and Fn more robust, and thus participating in subsequent feature calculation processing.

[0104] In this optional embodiment, the model parameters of the first initial encoding network are optimized using a comparison network. The internal structure of the comparison network is not limited. Figure 2dThe diagram illustrates the internal structure of a contrastive network. The contrastive network comprises a target contrastive module and a feature contrastive module.

[0105] The sample comparison module includes a feature interactions layer. This layer primarily captures complex relationships or interaction effects between input features, significantly improving the model's predictive performance, especially when the data has complex structures or nonlinear relationships. The sample candidate product features, positive behavior sequence features, and negative behavior sequence features output from the first initial encoding network can be input into the sample comparison module. In this module, the feature interactions layer performs feature interactions on the sample candidate product features, positive behavior sequence features, and negative behavior sequence features, respectively, resulting in updated versions. An alignment loss function is then constructed based on these updated features. Detailed implementation of the alignment loss function can be found in the aforementioned embodiments and will not be repeated here. Figure 2d In the diagram, the updated sample candidate product features, sample positive behavior sequence features, and sample negative behavior sequence features are represented as Ft, Fp, and Fn, respectively.

[0106] Among them, such as Figure 2d As shown, the feature comparison module includes a perturbation layer, a feature interactions layer, and a projector layer. The perturbation layer introduces controllable noise or variation to enhance the model's robustness and generalization ability. By perturbing the input features during training, the model can better adapt to inputs under different conditions and reduce overfitting to specific patterns. The projector layer transforms the input features into a vector representation with a specific dimension. This typically involves a series of linear or nonlinear transformations to make the output features more compact and expressive. A detailed description of the feature interactions module can be found in the aforementioned embodiments and will not be repeated here.

[0107] For example, the sample candidate product features, sample positive behavior sequence features, and sample negative behavior sequence features output by the first initial encoding network are input to the feature comparison module. In the feature comparison module, the sample positive behavior sequence features and sample negative behavior sequence features are augmented using a perturbation layer to obtain pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups. The sequences in the pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups are interacted with using a feature interaction layer to obtain updated pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups. According to the transformation layer, the updated pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups are vectorized to obtain the vectorized results of the sequences in the pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups; based on the vectorized results of the sequences in the pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups, a contrastive loss function is constructed.

[0108] Scenario-based implementation examples:

[0109] The following is an example, taking a shopping cart page as the target page, where the first positive action is to select the first item to be ordered in the shopping cart, the second positive action is to place an order, the negative action is to cancel the order, and the second item is the item to be canceled.

[0110] If the first product information participates in the cross-store discount activity, taking the discount of 40 yuan off for purchases over 300 yuan as an example, there are two first product information items, namely first product information S1 and first product information S2. The price of S1 is 120 yuan and the price of S2 is 150 yuan. The total price of S1 and S2 is 120 + 150 = 270 yuan. 300 - 270 = 30 yuan, which does not meet the discount threshold, so a third product information item needs to be recommended to the user.

[0111] Therefore, positive behavior sequence information can be obtained, including the user's selection behavior in the shopping cart, and the selected first product information S1 and first product information S2. Negative behavior sequence information can be obtained, namely the user's order cancellation behavior within a specified period (e.g., the previous 3 months), and the cancelled second product information T1 and second product information T2, etc. In addition, multiple candidate product information can be obtained, filtered from the product database based on the user's browsing history, purchase behavior, collection preferences, and other information, as well as the product's own attributes (e.g., category, brand, price range), popularity trends, seasonal factors, and other multi-dimensional information.

[0112] Positive and negative behavior sequence information is input into the target prediction model to predict the user's interest in performing a second positive behavior action on multiple candidate product information. Based on this interest, a subset of candidate products is selected as add-on items for the user. For example, a predetermined number of candidate products with high interest are selected as add-on items, or a subset with interest exceeding a predetermined interest threshold are selected.

[0113] For example, by displaying a pop-up window on the shopping cart page, the user can select a third item (Y1) from the list of add-on items. The price of Y1 is 31 yuan. In response to the user's selection, Y1 is added to the shopping cart and is then selected. This selection can be either default or manually marked by the user. The total price of the first item (S1), the second item (S2), and the third item (Y1) is 120 + 150 + 31 = 301 yuan, reaching the minimum purchase amount for a discount. The user then performs a second positive action (e.g., places an order) on these items, with an order amount of 301 - 40 = 261 yuan. In response to the user's order, order information is generated and submitted to the order system for processing.

[0114] It should be noted that if the user's first selected product does not meet the minimum purchase requirement for discounts, a "buy more, save more" control will be displayed on the target page. After the user triggers this control, add-on items will be displayed. These add-on items can be sorted by interest level from highest to lowest. They can be displayed on a single page or across multiple tabs, with different tabs displaying different add-on items. These add-on items can be differentiated by attributes such as price and category. For example, add-on items can be displayed across multiple tabs based on price ranges, with different tabs displaying different price ranges, such as 0-30 yuan, 30-80 yuan, 80-120 yuan, etc., each range sorted from highest to lowest interest. Alternatively, add-on items can be displayed across multiple tabs based on product category, with different tabs displaying different product categories, and within the same category, add-on items displayed from highest to lowest interest.

[0115] Optionally, the system can also count the number of times negative actions are performed on candidate product information, and de-weight the candidate product information based on the number of negative actions performed, thereby reducing the interest level of the candidate product information. The de-weighting method is not limited. For example, a linear de-weighting method can be used, reducing interest level based on the number of negative actions performed. For instance, the more times the negative actions are performed, the greater the decrease in interest level. Alternatively, a non-linear de-weighting method can be used, where the rate of decrease in interest level accelerates as the number of negative actions performed increases.

[0116] In addition to providing an information recommendation method, this application also provides a model training method. The process of the model training method provided in this application is described below.

[0117] Figure 3 A flowchart illustrating a model training method provided for an exemplary embodiment of this application is shown below. Figure 3 As shown, the method includes:

[0118] 301. Obtain multiple sample information, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information and their corresponding sample interest degree and sample success rate of performing the second positive behavior operation; wherein, the positive behavior sequence information is generated by performing the first positive behavior operation on the first product information, and the negative behavior sequence information is generated by performing the negative behavior operation on the second product information;

[0119] 302. Based on the positive and negative behavior sequence information of the samples, call the initial prediction model and the initial task constraint model to predict the interest and success rate of multiple sample candidate product information, and obtain the predicted interest and prediction success rate of multiple sample candidate product information.

[0120] 303. Construct a prediction loss function and a constraint loss function based on the predicted interest, sample interest, prediction success rate, and sample success rate corresponding to multiple sample candidate product information;

[0121] 304. Based on the prediction loss function and the constraint loss function, the initial prediction model and the initial task constraint model are jointly trained to obtain the target prediction model and the target task constraint model.

[0122] Regarding the embodiments provided in this application Figure 3 The detailed implementation methods and beneficial effects of each step in the method shown have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0123] An exemplary embodiment of this application provides a flowchart of an information prediction model, which includes an encoding network, an attention network, and a prediction network.

[0124] An encoding network is used to encode positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, short-term behavior sequence information, and multiple candidate product information to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features.

[0125] Attention networks are used to perform attention calculations on positive and negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features to generate attention features corresponding to multiple candidate product information.

[0126] The prediction network is used to predict the interest level of the attention features corresponding to multiple candidate product information, so as to obtain the interest level of multiple candidate product information to perform the second positive behavior operation.

[0127] The detailed implementation methods and beneficial effects of the above information recommendation model have been described in detail in the foregoing embodiments, and will not be elaborated upon here.

[0128] It should be noted that the execution subject of each step of the method provided in the above embodiments can be the same device, or the method can be executed by different devices. For example, the execution subject of steps 101 to 103 can be device A; or the execution subject of steps 101 and 102 can be device A, and the execution subject of step 103 can be device B; and so on.

[0129] Furthermore, some processes described in the above embodiments and accompanying drawings include multiple operations appearing in a specific order. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or they may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Additionally, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0130] Figure 4 A schematic diagram of the structure of an information recommendation device provided for an exemplary embodiment of this application, as shown below. Figure 4 As shown, the device includes: a display module 41, an acquisition module 42, a call module 43, and a recommendation module 44.

[0131] The display module 41 is used to display the target page, which includes the first product information of the target user performing the first positive behavior operation, and the target user performing the first positive behavior operation on the first product information to generate positive behavior sequence information, and the first positive behavior operation is associated with the second positive behavior operation.

[0132] The acquisition module 42 is used to acquire negative behavior sequence information generated by the target user performing negative behavior operations on the second product information within a specified time period if the first product information does not meet the set recommendation conditions;

[0133] Module 43 is invoked to predict the target user's interest in performing a second positive action on multiple candidate product information based on the positive action sequence information and the negative action sequence information.

[0134] The recommendation module 44 is used to recommend third product information based on the target user's interest in performing a second positive action on multiple candidate product information, so that the target user can jointly perform a second positive action on the first product information and the third product information.

[0135] In one optional embodiment, the recommendation module is specifically configured to: select target product information from multiple candidate product information based on the target user's interest in performing a second positive behavior operation on multiple candidate product information; display the target product information; and, in response to the operation of selecting a third product information from the target product information, display the third product information on the target page so that the target user can jointly perform a second positive behavior operation on the first product information and the third product information.

[0136] In an optional embodiment, the acquisition module is further configured to: acquire long-term behavioral sequence information in the first time dimension and short-term behavioral sequence information in the second time dimension corresponding to the target user, wherein the first time dimension is greater than the second time dimension; the invocation module is specifically configured to: invoke the target prediction model to predict the target user's interest in performing a second positive behavioral operation on multiple candidate product information based on the positive behavioral sequence information, negative behavioral sequence information, long-term behavioral sequence information and short-term behavioral sequence information.

[0137] Optionally, the calling module is specifically used to: encode features of positive behavior sequence information, negative behavior sequence information, long-term behavior sequence information, short-term behavior sequence information, and multiple candidate product information according to the target encoding network of the target prediction model, so as to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features; calculate attention on positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features according to the target attention network of the target prediction model, so as to generate attention features corresponding to multiple candidate product information; and predict interest on the attention features corresponding to multiple candidate product information according to the target prediction network of the target prediction model, so as to obtain the interest of multiple candidate product information in performing the second positive behavior operation.

[0138] Further optionally, the calling module is specifically used to: input the candidate product feature, positive behavior sequence feature, negative behavior sequence feature, long-term behavior sequence feature, and short-term behavior sequence feature into the target attention network for any candidate product feature; perform multi-head attention calculation on the candidate product feature with the positive behavior sequence feature and the negative behavior sequence feature respectively to obtain the first attention feature and the second attention feature corresponding to the candidate product information; and generate the third attention feature corresponding to the candidate product information based on the multi-head attention results between the candidate product feature and the long-term behavior sequence feature, and between the candidate product feature and the short-term behavior sequence feature.

[0139] Further optionally, the calling module is specifically used to: select at least a portion of the features that are compatible with the candidate product features from the long-term behavior sequence features as the updated long-term behavior sequence features; perform self-attention calculation on the short-term behavior sequence features to obtain the self-attention result; and generate a third attention feature corresponding to the candidate product information based on the multi-head attention result between the updated long-term behavior sequence features and the candidate product features, and the multi-head attention result between the candidate product features and the self-attention result.

[0140] Further optionally, the calling module is specifically used to: for any candidate product feature, generate a concatenated feature corresponding to the candidate product feature based on the first attention feature, the second attention feature, and the third attention feature; and perform feature prediction on the concatenated feature to obtain the interest degree corresponding to the candidate product feature.

[0141] In an optional embodiment, the system further includes a first training module. The first training module is configured to acquire multiple sample information, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information and their corresponding sample interest and success rate of performing the second positive behavior operation; based on the positive and negative behavior sequence information, it calls an initial prediction model and an initial task constraint model to predict the interest and success rate of the multiple candidate product information, obtaining the predicted interest and prediction success rate corresponding to the multiple candidate product information; based on the predicted interest, sample interest, prediction success rate, and sample success rate corresponding to the multiple candidate product information, it constructs a prediction loss function and a constraint loss function; based on the prediction loss function and constraint loss function, it jointly trains the initial prediction model and the initial task constraint model to obtain a target prediction model and a target task constraint model.

[0142] Optionally, the first training module is specifically used to: based on the positive and negative behavioral sequence information of the samples, call the initial prediction model to predict the interest level of multiple candidate product information; wherein, the initial prediction model also generates multiple sample migration features corresponding to the multiple candidate product information; based on the multiple sample migration features, the positive and negative behavioral sequence information of the samples, call the initial task constraint model to predict the success rate of the multiple candidate product information, and obtain the prediction success rate corresponding to each of the multiple candidate product information.

[0143] Optionally, the device further includes a second training module, which is further configured to: construct an alignment loss function for any sample candidate product feature based on the sample candidate product feature, sample positive behavior sequence feature, and sample negative behavior sequence feature output by the initial prediction network of the initial prediction model; perform data augmentation on the sample positive behavior sequence feature and sample negative behavior sequence feature respectively to obtain a pseudo-positive behavior sequence feature group and a pseudo-negative behavior sequence feature group; construct a contrastive loss function based on the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group; and iteratively train the initial encoding network based on the alignment loss function and the contrastive loss function to obtain the target encoding network.

[0144] Further optionally, the second training module is specifically used to construct a contrastive loss function with the goal of reducing the intra-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group, and increasing the inter-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group.

[0145] Regarding the embodiments provided in this application Figure 4 The detailed implementation methods and beneficial effects of each step in the apparatus shown have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0146] This application embodiment also provides a model training device, which includes:

[0147] The acquisition module is used to acquire multiple sample information, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information and their corresponding sample interest degree and the success rate of the sample executing the second positive behavior operation; wherein, the positive behavior sequence information is generated by executing the first positive behavior operation on the first product information, and the negative behavior sequence information is generated by executing the negative behavior operation on the second product information;

[0148] The calling module is used to call the initial prediction model and the initial task constraint model based on the positive behavior sequence information and negative behavior sequence information of the samples, and to predict the interest and success rate of multiple sample candidate product information, so as to obtain the predicted interest and prediction success rate of multiple sample candidate product information.

[0149] The construction module is used to construct the prediction loss function and the constraint loss function based on the prediction interest, sample interest, prediction success rate and sample success rate corresponding to multiple sample candidate product information.

[0150] The training module is used to jointly train the initial prediction model and the initial task constraint model based on the prediction loss function and the constraint loss function, so as to obtain the target prediction model and the target task constraint model.

[0151] The detailed implementation methods and beneficial effects of each step in the above-described device have been described in detail in the foregoing embodiments, and will not be elaborated upon here.

[0152] Figure 5 A schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application, such as... Figure 5 As shown, the device includes a memory 54 and a processor 55.

[0153] Memory 54 is used to store computer programs and can be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device.

[0154] The processor 55, coupled to the memory 54, is used to execute the computer program in the memory 54 for: executing the information recommendation method and model training method described in the foregoing embodiments, for details of which will not be repeated here.

[0155] Regarding the embodiments provided in this application Figure 5The detailed implementation methods and beneficial effects of each step in the device shown have been described in detail in the foregoing embodiments, and will not be elaborated here.

[0156] Furthermore, such as Figure 5 As shown, the electronic device also includes other components such as a communication component 56, a display 57, a power supply component 58, and an audio component 59. Figure 5 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 5 The components shown. Additionally... Figure 5 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the electronic device. The electronic device in this embodiment can be a desktop computer, laptop computer, smartphone, IoT (Internet of Things) device, or smart wearable device (such as a smartwatch or smart bracelet), or a server-side device such as a conventional server, cloud server, or server array. If the electronic device in this embodiment is a desktop computer, laptop computer, smartphone, or other terminal device, it may include... Figure 5 The components within the dashed box; if the electronic device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., it may be omitted. Figure 5 The component within the dashed box.

[0157] The aforementioned memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0158] The aforementioned communication component is configured to facilitate wired or wireless communication between the device containing the communication component and other devices. The device containing the communication component can access wireless networks based on communication standards, such as 2G, 3G, 4G / LTE, 5G, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0159] The aforementioned display includes a screen, which may include a Liquid Crystal Display (LCD) and a Touch Panel (TP). If the screen includes a Touch Panel, the screen can be implemented as a touchscreen to receive input signals from the user. The Touch Panel includes one or more touch sensors to sense touches, swipes, and gestures on the Touch Panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0160] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0161] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0162] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium.

[0163] Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0164] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0165] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An information recommendation method, characterized in that, include: Display a target page, the target page including first product information for which a target user performs a first positive action, and the target user generates a positive action sequence information by performing a first positive action on the first product information, and the first positive action is associated with a second positive action. If the first product information meets the set recommendation conditions, obtain the negative behavior sequence information generated by the target user performing negative behavior operations on the second product information within a specified time period; Based on the positive behavior sequence information and the negative behavior sequence information, the target prediction model is invoked to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information; Based on the target user's interest in performing a second positive behavioral operation on the multiple candidate product information, the target product information is selected from the multiple candidate product information; The target product information is displayed, and in response to the operation of selecting third product information from the target product information, the third product information is displayed on the target page so that the target user can jointly perform the second positive behavior operation on the first product information and the third product information.

2. The method according to claim 1, characterized in that, Also includes: Obtain long-term behavioral sequence information in a first time dimension and short-term behavioral sequence information in a second time dimension corresponding to the target user, wherein the first time dimension is greater than the second time dimension; Based on the positive behavior sequence information and the negative behavior sequence information, a target prediction model is invoked to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information, including: Based on the positive behavior sequence information, the negative behavior sequence information, the long-term behavior sequence information, and the short-term behavior sequence information, the target prediction model is invoked to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information.

3. The method according to claim 2, characterized in that, Based on the positive behavior sequence information, the negative behavior sequence information, and the behavior sequence information in the time dimension, the target prediction model is invoked to predict the target user's interest in performing a second positive behavior operation on multiple candidate product information, including: Based on the target encoding network of the target prediction model, feature encoding is performed on the positive behavior sequence information, the negative behavior sequence information, the long-term behavior sequence information, the short-term behavior sequence information, and multiple candidate product information to generate positive behavior sequence features, negative behavior sequence features, long-term behavior sequence features, short-term behavior sequence features, and multiple candidate product features. Based on the target attention network of the target prediction model, attention calculation is performed on the positive behavior sequence features, the negative behavior sequence features, the long-term behavior sequence features, the short-term behavior sequence features, and multiple candidate product features to generate attention features corresponding to the multiple candidate product information. Based on the target prediction network of the target prediction model, interest degree prediction is performed on the attention features corresponding to the multiple candidate product information to obtain the interest degree of the multiple candidate product information to perform the second positive behavior operation.

4. The method according to claim 3, characterized in that, Based on the target attention network of the target prediction model, attention calculations are performed on the positive behavior sequence features, the negative behavior sequence features, the long-term behavior sequence features, the short-term behavior sequence features, and multiple candidate product features to generate attention features corresponding to the multiple candidate product information, including: For any candidate product feature, the candidate product feature, the positive behavior sequence feature, the negative behavior sequence feature, the long-term behavior sequence feature, and the short-term behavior sequence feature are input into the target attention network; The candidate product features are subjected to multi-head attention calculation with the positive behavior sequence features and the negative behavior sequence features, respectively, to obtain the first attention feature and the second attention feature corresponding to the candidate product information; Based on the multi-head attention results between the candidate product features and the long-term behavior sequence features, and the multi-head attention results between the candidate product features and the short-term behavior sequence features, a third attention feature corresponding to the candidate product information is generated.

5. The method according to claim 4, characterized in that, Based on the multi-head attention results between the candidate product features and the long-term behavioral sequence features, and the multi-head attention results between the candidate product features and the short-term behavioral sequence features, a third attention feature corresponding to the candidate product information is generated, including: At least some features that match the candidate product features are selected from the long-term behavior sequence features as updated long-term behavior sequence features; Self-attention calculation is performed on the short-term behavioral sequence features to obtain the self-attention result; Based on the multi-head attention result between the updated long-term behavioral sequence features and the candidate product features, and the multi-head attention result between the candidate product features and the self-attention result, a third attention feature corresponding to the candidate product information is generated.

6. The method according to claim 4, characterized in that, Based on the target prediction network of the target prediction model, interest prediction is performed on the attention features corresponding to the multiple candidate product information to obtain the interest level of the multiple candidate product information in performing the second positive behavior operation, including: For any candidate product feature, a spliced ​​feature corresponding to the candidate product feature is generated based on the first attention feature, the second attention feature, and the third attention feature; Feature prediction is performed on the spliced ​​features to obtain the interest level corresponding to the candidate product features.

7. The method according to any one of claims 1-6, characterized in that, Also includes: Acquire multiple sample information, including: sample positive behavior sequence information, sample negative behavior sequence information, multiple sample candidate product information and their corresponding sample interest degree and sample success rate of performing the second positive behavior operation; Based on the positive behavior sequence information and the negative behavior sequence information of the samples, the initial prediction model and the initial task constraint model are invoked to predict the interest and success rate of the multiple sample candidate product information, thereby obtaining the predicted interest and prediction success rate corresponding to the multiple sample candidate product information. Based on the predicted interest degree corresponding to the multiple sample candidate product information, the sample interest degree, the prediction success rate, and the sample success rate, a prediction loss function and a constraint loss function are constructed. Based on the prediction loss function and the constraint loss function, the initial prediction model and the initial task constraint model are jointly trained to obtain the target prediction model and the target task constraint model.

8. The method according to claim 7, characterized in that, Based on the positive and negative behavior sequence information of the samples, the initial prediction model and the initial task constraint model are invoked to predict the interest and success rate of the multiple candidate product information samples, obtaining the predicted interest and prediction success rate corresponding to each of the multiple candidate product information samples, including: Based on the positive and negative behavioral sequence information of the samples, an initial prediction model is invoked to predict the interest level of the multiple candidate product information samples; wherein, the initial prediction model also generates multiple sample migration features corresponding to the multiple candidate product information samples. Based on the multiple sample migration features, positive sample behavior sequence information, and negative sample behavior sequence information, the initial task constraint model is invoked to predict the success rate of the multiple sample candidate product information, thereby obtaining the prediction success rate corresponding to each of the multiple sample candidate product information.

9. The method according to claim 7, characterized in that, Also includes: For any sample candidate product feature, an alignment loss function is constructed based on the sample candidate product features, sample positive behavior sequence features, and sample negative behavior sequence features output by the initial prediction network of the initial prediction model; Data augmentation is performed on the positive behavior sequence features and negative behavior sequence features of the samples to obtain pseudo-positive behavior sequence feature groups and pseudo-negative behavior sequence feature groups; Based on the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group, a contrastive loss function is constructed; The initial encoding network of the initial prediction model is iteratively trained based on the alignment loss function and the contrast loss function to obtain the target encoding network of the target prediction model.

10. The method according to claim 9, characterized in that, Based on the pseudo-positive behavior sequence feature set and the pseudo-negative behavior sequence feature set, a contrastive loss function is constructed, including: A contrastive loss function is constructed with the goal of reducing the intra-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group, and increasing the inter-group distance between the pseudo-positive behavior sequence feature group and the pseudo-negative behavior sequence feature group.

11. A model training method, characterized in that, include: Multiple sample information is acquired, including: positive behavior sequence information, negative behavior sequence information, multiple candidate product information and their corresponding sample interest degree and sample success rate of performing the second positive behavior operation; wherein, the positive behavior sequence information is generated by performing the first positive behavior operation on the first product information, and the negative behavior sequence information is generated by performing the negative behavior operation on the second product information; Based on the positive behavior sequence information and the negative behavior sequence information of the samples, the initial prediction model and the initial task constraint model are invoked to predict the interest and success rate of the multiple sample candidate product information, thereby obtaining the predicted interest and prediction success rate corresponding to the multiple sample candidate product information. Based on the predicted interest degree corresponding to the multiple sample candidate product information, the sample interest degree, the prediction success rate, and the sample success rate, a prediction loss function and a constraint loss function are constructed. Based on the prediction loss function and the constraint loss function, the initial prediction model and the initial task constraint model are jointly trained to obtain the target prediction model and the target task constraint model.

12. An electronic device, characterized in that, include: Memory and processor; The memory is used to store a computer program; the processor, coupled to the memory, is used to execute the computer program to implement the steps of the method according to any one of claims 1-10 and 11.

13. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it causes the processor to perform the steps of the method according to any one of claims 1-10 and 11.

14. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the processor is caused to perform the steps in the method of any one of claims 1-10 and 11.

Images