Interaction behavior effective data prediction model training method and device, equipment and medium

Abstract

The embodiment of the application provides an interactive behavior effective data estimation model training method, device, equipment and medium, which is used for estimating the interactive behavior effective data of a user and a merchant. In the method, sample data used for training the interactive behavior effective data estimation model is obtained, including sample input data and sample output data. The sample input data includes identification information of a sample merchant and text semantic description information of the sample merchant, and the sample output data is sample interactive behavior effective data. Through the model processing, a semantic alignment loss value between the identification information of the sample merchant and the text semantic description information of the sample merchant is obtained, an interactive behavior effective data estimation loss value between the estimated interactive behavior effective data and the sample interactive behavior effective data is estimated, an entropy regular loss value corresponding to the estimated interactive behavior effective data is estimated, and a group-level gradient gating weight is estimated, so that the model is adjusted in parameters. The accuracy of the interactive behavior effective data between the user and the merchant is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method, apparatus, electronic device, and computer storage medium for training an effective data prediction model for interactive behavior. This application also relates to a method, apparatus, electronic device, and computer storage medium for obtaining effective data on interactive behavior. Background Technology

[0002] With the rapid development of online-to-offline service platforms, personalized recommendation advertising systems are widely used. Existing personalized recommendation advertising systems generally use historical user-merchant interactions (e.g., clicks, browsing, favorites, adding to cart, placing orders, etc.) as monitoring signals, training a model to predict the effectiveness of these interaction data. This trained model then analyzes and predicts the probability of a valid interaction between the user and the merchant, such as the probability of a user successfully placing an order for a meal offered by that merchant.

[0003] However, in the training process of the existing technology, all user-merchant interactions are generally treated as valid positive samples for model training. This results in the model learning a large number of low-efficiency or even invalid supervision signals during parameter optimization, which misleads the model in understanding the user's true intention and causes the following defects: for merchants with high click-through rates but low effective interactions, the model's estimated probability of effective interaction between the user and the merchant deviates significantly from the actual probability.

[0004] Therefore, how to solve the problem that the probability of effective interaction between users and merchants predicted by the models provided in the existing technology is too low, resulting in a poor user experience. Summary of the Invention

[0005] This application provides a method for training an effective data prediction model for interactive behavior, thereby improving the probability of effective user-merchant interactions predicted by the model and thus enhancing the user experience. This application also provides a method, apparatus, electronic device, and computer storage medium for obtaining effective interactive behavior data.

[0006] The specific plan is as follows: In a first aspect, embodiments of this application provide a method for training an effective data prediction model for interactive behavior. The effective data prediction model is used to predict effective data of user-merchant interaction behavior. The method includes: acquiring sample data for training the effective data prediction model, the sample data including sample input data and sample output data, the sample input data including the identifier information of the sample merchant and the text semantic description information of the sample merchant, and the sample output data being sample effective data of interactive behavior expected to be obtained based on the sample input data to represent the effectiveness of the interactive behavior; providing the sample input data to an initial effective data prediction model for interactive behavior, obtaining the predicted effective data of interactive behavior output by the initial effective data prediction model, and obtaining a semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial effective data prediction model; obtaining the effective data prediction loss value of interactive behavior between the predicted effective data of interactive behavior output by the initial effective data prediction model and the sample effective data of interactive behavior; and according to... The estimated effective data of the interaction behavior and the weights corresponding to the estimated effective data of the interaction behavior are used to obtain the entropy regularization loss value corresponding to the estimated effective data of the interaction behavior; the sample real feature data in the sample input data are obtained, and the group-level gradient gating weights of the activated sample feature data corresponding to the sample input data are obtained according to the positive weights of the negative impact of the sample real feature data on the order fulfillment success result. The group-level gradient gating weights are the negative weights of the activated sample feature data on the negative impact of the order fulfillment success result. The order corresponding to the order fulfillment success result is the order generated based on the interaction behavior between the user and the merchant. The activated sample feature data are the sample feature data selected from the sample feature data corresponding to the sample input data to obtain the estimated effective data of the interaction behavior and the semantic alignment loss value; the initial effective data of the interaction behavior prediction model is adjusted according to the estimated loss value of the effective data of the interaction behavior, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights to obtain the trained effective data of the interaction behavior prediction model.

[0007] Secondly, embodiments of this application provide a training apparatus for an effective data prediction model of interactive behavior. The effective data prediction model of interactive behavior is used to predict effective data of user-merchant interaction behavior. The apparatus includes: an acquisition unit, used to acquire sample data for training the effective data prediction model of interactive behavior; the sample data includes sample input data and sample output data; the sample input data includes the identifier information of the sample merchant and the text semantic description information of the sample merchant; the sample output data is sample effective data of interactive behavior expected to be obtained based on the sample input data, representing the effectiveness of the interactive behavior; a model analysis unit, used to provide the sample input data to an initial effective data prediction model of interactive behavior, obtain the predicted effective data of interactive behavior output by the initial effective data prediction model of interactive behavior, and obtain the semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial effective data prediction model of interactive behavior; and an effective data prediction loss value acquisition unit, used to obtain the effective data prediction loss value of interactive behavior between the predicted effective data of interactive behavior output by the initial effective data prediction model of interactive behavior and the sample effective data of interactive behavior. The regularization loss value acquisition unit is used to obtain the entropy regularization loss value corresponding to the estimated interaction behavior effective data based on the estimated interaction behavior effective data and the corresponding weights of the estimated interaction behavior effective data. The group-level gradient gating weight acquisition unit is used to acquire the sample real-valued feature data in the sample input data, and obtain the group-level gradient gating weights of the activated sample feature data corresponding to the sample input data based on the positive weights of the negative impact of the sample real-valued feature data on the order fulfillment success result. The group-level gradient gating weights are the weights of the activated sample feature data on the order fulfillment success result. The negative impact of the successful order fulfillment result is weighted inversely. The order corresponding to the successful order fulfillment result is an order generated based on the interaction behavior between the user and the merchant. The activated sample feature data is the sample feature data selected from the sample input data to obtain the estimated effective interaction behavior data and the semantic alignment loss value. The training unit is used to adjust the initial effective interaction behavior data prediction model based on the estimated loss value of the effective interaction behavior data, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights to obtain the trained effective interaction behavior data prediction model.

[0008] Thirdly, embodiments of this application provide a method for obtaining effective interactive behavior data. The method includes: obtaining service information required by a user and candidate merchant information that can provide the service information; providing the service information required by the user and the candidate merchant information as input data to a trained effective interactive behavior data prediction model to obtain effective interactive behavior data between the user and the candidate merchants output by the effective interactive behavior data prediction model, wherein the effective interactive behavior data prediction model is trained using the training method described in the first aspect; and sorting multiple candidate merchant information according to the effective interactive behavior data between the user and the candidate merchants to obtain sorting information corresponding to each of the multiple candidate merchant information.

[0009] Fourthly, embodiments of this application provide an apparatus for obtaining effective interactive behavior data. The apparatus includes: an information acquisition unit, configured to acquire service information required by a user and candidate merchant information that can provide the service information; an effective interactive behavior data acquisition unit, configured to provide the service information required by the user and the candidate merchant information as input data to a trained effective interactive behavior data prediction model, thereby obtaining effective interactive behavior data between the user and the candidate merchants output by the effective interactive behavior data prediction model, wherein the effective interactive behavior data prediction model is trained using the training method described in the first aspect; and a sorting unit, configured to sort multiple candidate merchant information according to the effective interactive behavior data between the user and the candidate merchants, thereby obtaining sorting information corresponding to each of the multiple candidate merchant information.

[0010] Fifthly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory is used to store one or more computer instructions; the processor is used to execute the one or more computer instructions to implement the method described in the first aspect.

[0011] In a sixth aspect, embodiments of this application provide a computer-readable storage medium having stored thereon one or more computer instructions that are executed by a processor to implement the method described in the first aspect.

[0012] Compared with the prior art, this application has the following advantages: This application provides a method for training an effective data prediction model for interactive behavior. The effective data prediction model is used to predict effective data of user-merchant interaction behavior. The method includes: acquiring sample data for training the effective data prediction model, the sample data including sample input data and sample output data. The sample input data includes the identifier information and textual semantic description information of the sample merchant. The sample output data is the sample effective data of interactive behavior expected to be obtained based on the sample input data, representing the effectiveness of the interactive behavior. The sample input data is provided to an initial effective data prediction model to obtain the predicted effective data of interactive behavior output by the model, and the semantic alignment loss value between the identifier information and the textual semantic description information of the sample merchant is obtained. The identifier information of the sample merchant is the identifier provided by the online-to-offline service platform to the sample merchant, and the textual semantic description information of the sample merchant is used to describe the services that the sample merchant can provide to users. The semantic alignment loss between identifier information and text semantic description information is calculated. Based on this loss, the identifier information or text semantic description information is continuously changed to obtain text semantic description information matching the identifier information of the sample merchant, or vice versa, establishing an information binding process between identifier information and text semantic description information. This allows the online-to-offline service platform to obtain not only the identifier information of the sample merchant but also the description information of the services that the sample merchant can provide to users, increasing the probability that the platform will recommend the sample merchant to users. Then, by providing user and merchant information to the model, the probability of the model analyzing the effectiveness of the interaction between the user and the merchant is improved.

[0013] After providing the sample input data to the initial interaction behavior effective data prediction model, the model's output predicted interaction behavior effective data is also obtained. The interaction behavior effective data prediction loss value between the predicted interaction behavior effective data and the sample interaction behavior effective data is calculated. This loss value quantifies the model's error, thus determining the parameters that need adjustment. Based on the predicted interaction behavior effective data and its corresponding weights, the entropy regularization loss value is obtained. The entropy regularization loss value represents the diversity of activated sample feature data in the sample input data. A larger entropy regularization loss value indicates lower diversity of activated sample feature data in the sample input data, and a higher analysis frequency for a particular activated sample feature data. A smaller entropy regularization loss value indicates higher diversity of activated sample feature data in the sample input data, meaning that the model analyzes a wider variety of activated sample feature data types in multiple training rounds, thereby improving the accuracy of the model's predicted interaction behavior effective data. Calculate the estimated loss value and entropy regularization loss value of the effective data of the interaction behavior, and determine the parameters that need to be adjusted by the model from the forward path of the model.

[0014] The process involves acquiring real-valued feature data from the input data. Based on the positive weights of these real-valued feature data in relation to the negative impact on order fulfillment, group-level gradient gating weights are obtained for the activated feature data corresponding to the input data. These group-level gradient gating weights represent the negative weights of the activated feature data in relation to the negative impact on successful order fulfillment. Here, group-level gradient gating weights are obtained from the reverse training path of the model to control the gradients of the parameters that need to be adjusted in the model, obtained from the estimated loss value and entropy regularization loss value of the effective data of the interaction behavior. Different gradient amounts are applied to different parts of the model to achieve different gradient adjustments for parameters in different parts of the model. For example, by using the estimated loss value and entropy regularization loss value of the effective data of the interaction behavior, positive sample feature data contributing to successful order fulfillment and negative sample feature data having a negative impact on successful order fulfillment are deduced. The gradient adjustment weights of the parameters of the modules processing positive sample feature data are increased, while the gradient adjustment weights of the parameters of the modules processing negative sample feature data are decreased.

[0015] Based on the above analysis, this method determines the parameters that the model needs to adjust by estimating the loss value using effective data of interactive behavior. Combined with group-level gradient gating weights, the update magnitude of the parameters to be adjusted is determined. Entropy regularization loss value is used to determine the diversity of activated analysis data in the sample input data. Semantic alignment loss value is used to improve the semantic transfer process between the merchant's identification information and textual semantic description information, enhancing the model's understanding of the services that the merchant can provide to the user. Therefore, the method provided in this application can improve the accuracy of the trained model in analyzing effective data of interactive behavior between the user and the sample merchant, thereby improving the user experience. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating the application scenario of the solution provided in the embodiments of this application.

[0017] Figure 2 A flowchart of the training method for the effective data prediction model of interactive behavior provided in the first embodiment of this application.

[0018] Figure 3 This is a schematic diagram illustrating the training process of the effective data prediction model for interactive behavior provided in the first embodiment of this application.

[0019] Figure 4 A schematic diagram of the training device for the effective data prediction model of interactive behavior provided in the second embodiment of this application.

[0020] Figure 5 A flowchart of a method for obtaining effective interactive behavior data provided in the third embodiment of this application.

[0021] Figure 6 This is a schematic diagram of an interactive behavior effective data acquisition device provided in the fourth embodiment of this application.

[0022] Figure 7 This is a schematic diagram of an electronic device provided in the fifth embodiment of this application. Detailed Implementation

[0023] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0024] It should be noted that the terms "first," "second," "third," etc., in the claims, specification, and drawings of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. Such data are interchangeable where appropriate so that the embodiments of this application described herein can be implemented in a sequence other than that shown or described herein. Furthermore, the terms "comprising," "having," and their variations are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0025] It should be understood that in the embodiments of this application, "at least one" means one or more, and "more than one" means two or more. "And / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the related objects before and after it are in an "or" relationship. "Contains A, B and / or C" means containing any one, two, or three of A, B, and C.

[0026] It should be understood that in the embodiments of this application, "B corresponding to A", "B corresponding to A", "A corresponds to B" or "B corresponds to A" means that B is associated with A, and B can be determined based on A. Determining B based on A does not mean that B is determined solely based on A; B can also be determined based on A and / or other information.

[0027] Based on the reasons mentioned in the background art, the first embodiment of this application provides a training method for an effective data prediction model of interactive behavior. The method determines the parameters that the model needs to adjust by using the effective data prediction loss value of interactive behavior, and determines the update magnitude of the parameters to be adjusted by combining group-level gradient gating weights. The entropy regularization loss value is used to determine the diversity of activated analysis data in the sample input data. The semantic alignment loss value is used to improve the semantic transfer process between the identification information and textual semantic description information of the sample merchant, thereby improving the model's understanding of the service information that the sample merchant can provide to the user. Therefore, the method provided in this application can improve the accuracy of the trained model in analyzing effective data of interactive behavior between the user and the sample merchant, thereby improving the user experience.

[0028] The prediction model obtained by the interaction behavior effective data prediction model training method provided in this application can be used in e-commerce platforms to recommend products to users. For example, it can display recommended products and a list of candidate merchants that can provide the products on the homepage of the target application. The list of candidate merchants that can provide the products can be used by the prediction model obtained through this training method to predict the effective data of interaction behavior between the user and each candidate merchant, such as predicting the click-through rate or the order rate. Based on the effective data of interaction behavior between the user and each candidate merchant, the ranking information of the candidate merchant information is obtained, and then the multiple candidate merchant information is ranked according to a preset ranking rule to obtain the candidate merchant information list shown above. Alternatively, it can be used to display a list of candidate merchants that can provide the target service information based on the user's query request for target service information in the target application. This application does not specifically limit the specific application scenarios of the trained interaction behavior effective data prediction model.

[0029] To understand the method embodiments of this application, their application scenarios are described. Please refer to... Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario of the solution provided in the embodiments of this application. This application scenario is merely an illustrative example and is not intended to limit the specific application scenario. Figure 1 As shown, in this application scenario, a server 102 and a client 101 are provided. In this embodiment, the client 101 and the server 102 establish a connection through network communication to transmit data.

[0030] Client 101 can be an electronic device with display and data processing capabilities, such as a mobile phone, tablet, smartwatch, desktop computer, smart TV, VR device, in-vehicle device, wearable device, or laptop. Client 101 is used to obtain service object information provided by merchants on the current page browsed by the user on the service platform. It can also obtain service object information input by the user, and user interaction information regarding the service object information. Based on the service object information and interaction information, it obtains the service information required by the user and provides this information to server 102. Server 102 then provides the user with a list of candidate merchants that can provide the required service information. This list includes the ranking information of each candidate merchant. The ranking information provided by server 102 can be obtained by using an effective data prediction model of interactive behavior to estimate the probability of effective interaction between the user and each candidate merchant.

[0031] Server 102 possesses high computing power. Server 102 can be a server with high-speed central processing unit (CPU) computing power, long-term reliable operation, strong input / output (I / O) external data throughput capacity, and better scalability. Server 102 can be a single server or a server cluster. Server 102 is used to provide candidate merchant information to client 101. Server 102 can be configured with a pre-trained interactive behavior effective data prediction model. It sends user information and candidate merchant information associated with the user's service requests to the trained interactive behavior effective data prediction model. This model analyzes the probability of valid data from user interactions with each merchant, thereby ranking the multiple candidate merchant information. Server 102 then sends the ranked candidate merchant information list to client 101. Server 102 also provides other specific services to client 101, such as website access and application access, which are not specifically limited in this application.

[0032] Client 101 and server 102 can communicate using various communication systems, such as wired or wireless communication systems. Wireless communication systems can include, for example, Global System for Mobile Communications (GSM), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), General Packet Radio Service (GPRS), Long Term Evolution (LTE), LTE Frequency Division Duplex (FDD), LTE Time Division Duplex (TDD), Universal Mobile Telecommunication System (UMTS), Worldwide Interoperability for Microwave Access (WiMAX), future 5th generation (5G) systems or new radio (NR), and satellite communication systems.

[0033] First Embodiment The first embodiment of this application provides a method for training an effective data prediction model for interactive behavior. This method is applied to electronic devices, such as servers, desktop computers, laptops, smart TVs, VR devices, in-vehicle devices, wearable devices, mobile phones, tablets, smartwatches, and other electronic devices with data processing functions.

[0034] Please refer to Figure 2 As shown, the first embodiment of this application provides a training method for an effective data prediction model of interactive behavior, wherein the effective data prediction model of interactive behavior is used to predict the effective data of interactive behavior between users and merchants, and the method includes the following steps S101 to S106.

[0035] Step S101: Obtain sample data for training the effective data prediction model of interactive behavior. The sample data includes sample input data and sample output data. The sample input data includes the identification information of the sample merchants and the textual semantic description information of the sample merchants. The sample output data is the sample interactive behavior effective data expected to be obtained based on the sample input data to represent the effectiveness of interactive behavior.

[0036] This step is used to obtain the sample data required for training the model, which will be used in subsequent steps to train the effective data prediction model for interaction behavior, in order to obtain a model for predicting effective data for user and merchant interaction behavior.

[0037] Valid user-merchant interaction data refers to interaction data where the user's request and the merchant's service satisfy semantic matching, and the system can fulfill the request. In other words, valid interaction data includes the user's true intent, and the merchant provided by the system can genuinely meet that need. For example, for the product the user needs, the merchant can currently provide the product and deliver it.

[0038] First, interactive behavior data of successful booking conversions.

[0039] Second, user interaction data with high intent. For example, interaction data such as viewing product delivery information and viewing merchant details.

[0040] Invalid interaction data refers to data related to user requests where the system cannot fulfill certain fulfillment conditions, resulting in unsuccessful fulfillment. For example, a user requests a desired service, but the merchant providing that service in the system is unable to provide it due to at least one of the following reasons: the service is unavailable, the current time is outside the merchant's service hours, or the delivery distance exceeds a preset distance threshold. Therefore, this type of interaction data lacks signal validity and is considered invalid data.

[0041] The sample input and output data provided in this step can be obtained from the system's historical logs, which record at least one complete data entry of the user-merchant interaction process.

[0042] The sample input data includes merchant-related input data and user-related input data.

[0043] The merchant identification information in the sample is provided by the system, such as the Merchant ID (MerchantIdentifier), which is a unique identifier for each merchant registered on the system. However, for newly registered merchants, the system is unaware of other descriptive information about them. Therefore, if a user submits a service request, the system cannot recommend the merchant because it lacks information about the specific services the merchant can provide, resulting in a low recommendation rate for that merchant.

[0044] Therefore, it is necessary to generate corresponding textual semantic descriptions for the service categories provided by the merchant. The merchant's identification information and textual semantic descriptions are then fused together so that the system can obtain both the merchant's identification information and the descriptions of the services the merchant can provide.

[0045] The textual semantic description information of the sample merchants includes their feature data, which includes basic feature data and contextual difference feature data between the sample merchants and similar merchants. This textual semantic description information can be obtained in the following ways:

[0046] The sample input data includes sample merchant feature data, which includes the merchant's identification information and textual semantic description information; the sample merchant feature data is obtained in the following way: Obtain the basic description information of the sample merchants; provide the basic description information of the sample merchants to the frozen large model to obtain the basic description semantic vector of the sample merchants and a candidate merchant feature vector library, wherein the candidate merchant feature vector library includes the basic description semantic vectors corresponding to multiple merchants generated by the frozen large model; based on the context information and constraints corresponding to the service categories that the sample merchants can provide to users, select a preset number of candidate merchants' basic description semantic vectors associated with the sample merchants from the candidate merchant feature vector library; based on the basic description semantic vector of the sample merchants and the basic description semantic vector of the candidate merchants, obtain the context constraint difference semantic vector between the sample merchants and the candidate merchants; use the basic description semantic vector of the sample merchants and the context constraint difference semantic vector as the feature data of the sample merchants.

[0047] The basic descriptive information of the sample merchants is natural description information. The frozen large model refers to a large model that has been trained and does not require parameter updates. It is used to extract features from the basic descriptive information of merchants and convert text into vectors.

[0048] Specifically, a semantic summary is generated based on the structured prompts. This semantic summary is then converted into a semantic vector, which represents the basic semantic description vector of the sample merchant. The frozen large model extracts features from the basic description information of the merchants to generate corresponding basic semantic description vectors. All the generated basic semantic description vectors of multiple candidate merchants are stored in the candidate merchant feature vector library.

[0049] After obtaining the basic semantic description vector (stext) of the sample merchants and the candidate merchant feature vector library provided by the frozen large model, information on candidate merchants that are related to the sample merchants is filtered from the candidate merchant feature vector library. This filtering can be based on the contextual information and constraints corresponding to the service categories provided by the sample merchants to users.

[0050] The contextual information corresponding to the service categories provided by the sample merchants to users can refer to the service scenario information corresponding to the current order provided by the sample merchants to users, such as time information, the user's geographical location, weather conditions, the user's real-time interactive behavior in the current scenario, and so on.

[0051] The constraints corresponding to the service categories provided by the sample merchants to users can be the conditions under which the services provided by the sample merchants to users can be executed, such as the minimum order price set by the merchant, the order delivery fee, and the geographical area where the merchant is located.

[0052] Based on the context and constraints, select candidate merchants from the candidate merchant feature vector library that meet at least the following conditions: The geographical region where the candidate merchant is located is the same as the geographical region where the sample merchant is located, or the geographical region where the candidate merchant is located is adjacent to the geographical region where the sample merchant is located. The price difference between the price of goods offered to users by candidate merchants and the price of goods offered to users by sample merchants is less than a preset price difference threshold; The difference between the delivery fee for candidate merchants to deliver goods to target users and the delivery fee for sample merchants to deliver goods to target users is less than the preset delivery fee difference threshold.

[0053] Based on the above context and constraints, a preset number of candidate merchants associated with the sample merchants are obtained as basic semantic vectors.

[0054] The context constraint difference semantic vector between the sample merchant and the candidate merchant refers to the differentiated service feature data that the sample merchant can provide to the user compared with a preset number of candidate merchants. It is used to represent the distinctive differentiated services that the sample merchant can provide to the user compared with the candidate merchants.

[0055] Specifically, the context constraint difference semantic vector between the sample merchant and the candidate merchant is obtained as follows: the basic description semantic vectors of the preset number of candidate merchants are subjected to equal pooling to obtain the equal pooled description semantic vector of the candidate merchant; the residual vector between the basic description semantic vector of the sample merchant and the equal pooled description semantic vector of the candidate merchant is obtained as the context constraint difference semantic vector between the basic description semantic vector of the sample merchant and the equal pooled description semantic vector.

[0056] The process of performing average pooling on the basic description semantic vectors of a preset number of candidate merchants refers to averaging the multiple basic feature attributes in the basic semantic vectors of the preset number of candidate merchants to obtain an average vector.

[0057] Using multiple basic characteristic attributes of merchants as screening conditions, three candidate merchants associated with sample merchant A are obtained: first candidate merchant B, second candidate merchant C, and third candidate merchant D.

[0058] The basic description semantic vectors of the above three candidate merchants are scored according to their multiple dimensional features, as shown in Table 1: The feature numbers corresponding to the above three dimensions are then averaged: First dimension feature: (0.7 + 0.6 + 0.8) / 3 = 0.7 Second dimension feature: (0.5 + 0.6 + 0.7) / 3 = 0.6 Third dimension feature: (1.0 + 0.5 + 1.0) / 3 = 0.83 Therefore, the basic descriptive semantic feature vectors of the above three candidate merchants are subjected to average pooling, and the resulting average pooled descriptive semantic vectors are [0.7, 0.6, 0.83], which represent the average value of the three dimensions of features of the three candidate merchants.

[0059] Among them, the basic description semantic vector of the above sample merchant A is [0.95, 0.9, 1.0].

[0060] Specifically, obtaining the residual vector between the basic description semantic vector of the sample merchant and the average pooling description semantic vector of the candidate merchant means: taking the difference between the basic description semantic vector of the sample merchant and the average pooling description semantic vector, that is, subtracting the feature value of the average pooling description semantic vector from the feature value of the basic description semantic vector of the sample merchant.

[0061] The specific calculation process is as follows: The feature difference for the first feature dimension is 0.95 - 0.7 = +0.25, indicating that the value of the first feature dimension of sample merchant A is greater than the average value. The feature difference for the second feature dimension is 0.9 - 0.6 = +0.3, indicating that the value of the second feature dimension of sample merchant A is greater than the average value. The feature difference for the third feature dimension is 1.0 - 0.83 = +0.17, indicating that the value of the third feature dimension of sample merchant A is greater than the average value. Therefore, the context-constrained differential semantic vector between the basic description semantic vector and the average pooling description semantic vector of the sample merchant is [0.25, 0.3, 0.17].

[0062] By obtaining the basic descriptive semantic vector and context-constrained differentiated semantic vector of the sample merchants, we can obtain their distinctive features. Therefore, when a user submits a target service request on an online-to-offline service platform, the platform will prioritize sample merchant A among the other three candidate merchants when providing them with candidate merchants.

[0063] Therefore, based on the above method, while adding basic semantic description vectors to the sample merchants, context constraint difference semantic vectors are also added. The basic semantic description vectors and context constraint difference semantic vectors are used as vectors corresponding to the text semantic description information of the sample merchants and stored in the context database.

[0064] Next, the sample input data is provided to the initial interaction behavior effective data prediction model for data processing. Because the sample merchants have relatively rich descriptive information, it is beneficial for the model to analyze the accuracy of the effective data of interaction behavior between sample merchants and sample users.

[0065] Step S102: Provide the sample input data to the initial interaction behavior effective data prediction model, obtain the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model, and obtain the semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial interaction behavior effective data prediction model.

[0066] This step is used to obtain the effective data of the predicted interaction behavior after the model analyzes the sample input data. The effective data of the predicted interaction behavior includes at least one of the following: predicted click-through rate data and predicted order rate data.

[0067] The initial interaction behavior effective data prediction model can be a multi-task model or a multi-expert model, which divides the sample input data into multiple different categories and processes the sample input data of different categories separately. Specifically: The sample input data includes sample input data of multiple categories; the step of providing the sample input data to the initial interaction behavior effective data prediction model to obtain the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model includes: providing the sample input data of the multiple categories to the initial interaction behavior effective data prediction model, and performing the following steps through the initial interaction behavior effective data prediction model: obtaining sample feature data corresponding to the sample input data of each category as sample feature data of each category; obtaining the first activated sample feature data in the sample feature data of each category; fusing the first activated sample feature data of all categories to obtain the first feature activation fusion result; obtaining the second activated sample feature data in the sample feature data of each category according to the first feature activation fusion result and the weight corresponding to the sample feature data of each category; fusing the second activated sample feature data of all categories to obtain the second feature activation fusion result; and obtaining the predicted interaction behavior effective data according to the second feature activation fusion result.

[0068] The effective data prediction model for initial interaction behavior includes forward and backward pathways, such as... Figure 3 Both the forward and backward pathways use the same sample input data. The forward pathway analyzes the sample input data to obtain effective data for predicting interaction behavior, thereby obtaining the predicted loss value of the effective data for interaction behavior in step S104. The backward pathway data obtains the sample real-valued feature data from the sample input data. By analyzing the weights of the abnormal impact of the sample real-valued feature data on the order fulfillment success result, the weights of the gradients that need to be adjusted to control the parameters of different parts of the model are obtained. The backward pathway data is described in detail in subsequent step S105.

[0069] In step S103, the forward path process begins first: The effective data prediction model for the initial interaction behavior is the two-layer multi-hybrid expert model provided in the embodiments of this application (such as...). Figure 3 As shown), it processes the sample input data as follows: The sample input data is classified according to categories to obtain sample input data of multiple categories. Among them, the multiple categories in this embodiment are as follows: Figure 3 As shown, there are merchant group, user group, context group, and constraint group.

[0070] In the first layer of the multi-hybrid expert model, sample feature data corresponding to the sample input data of each expert group is obtained. This refers to the sample feature data obtained by each expert group extracting the original feature data from the sample input data. The sample feature data corresponding to the four expert groups are provided to the sharing group to obtain the sample feature data corresponding to the sharing group. The sample feature data for the merchant group consists of feature data associated with the service categories that merchants can provide to users. The sample feature data for the user group consists of feature data representing the degree of user interest in service categories. The sample feature data for the context group consists of service scenario feature data involved in the current service process for the service request provided by the user, such as time features, location features, and weather features of the current service scenario. The sample feature data for constraints consists of feature data indicating whether the merchant can provide the service for the user's service request. The sample feature data for the sharing group includes: common feature data obtained after cross-group interaction of the above four groups of sample feature data.

[0071] Each expert group belongs to an expert sub-network and processes the sample feature data within that expert group. Specifically, a preset number of sample feature data are selected for initial activation within each expert group; the remaining inactive sample feature data are not included in the analysis and are not used to calculate the fusion result. Figure 3 In the first-layer multi-hybrid expert model, within each expert group, the first feature state 31 indicates that the sample feature data belongs to the first activated sample feature data, and the second feature state 32 indicates that the sample feature data belongs to the first unactivated sample feature data. The selection of a preset number of first activated sample feature data in each expert group is based on the semantic matching degree between each sample feature data and the current sample input data. Based on the semantic matching degree corresponding to each sample feature data, sample feature data whose ranking position in this expert group is within a preset position (e.g., the first two positions) is obtained as the first activated sample feature data.

[0072] For example, the current sample input data is: users like food with a specified flavor and like to enjoy that food in a specified service scenario. Merchants are restaurants that can provide the specified flavor and can provide the first service scenario.

[0073] Calculate the semantic matching degree between the sample feature data and the sample input data in each of the five expert groups. Select the first two sample feature data as the first activated sample feature data in each expert group based on the semantic matching degree.

[0074] The first activated sample feature data of all expert groups in the first-layer multi-hybrid expert model are fused to obtain the first feature activation fusion result. Specifically, the first fusion weight corresponding to each expert group in the first-layer multi-hybrid expert model is obtained. Based on the first fusion weight corresponding to each expert group and the first activated sample feature data corresponding to that expert group, the first feature activation fusion result of the first-layer multi-hybrid expert model is obtained. Among them, the first fusion weight G1 corresponding to each expert group is obtained by back-calculation from the interaction behavior effective data prediction loss value obtained by the initial interaction behavior effective data prediction model, and is used to represent the importance of the first activated sample feature data of that expert group to the first feature activation fusion result.

[0075] The following table (Table 2) lists the first activated sample feature data, the semantic matching degree corresponding to the first activated sample feature data, and the first fusion weight corresponding to each expert group in the first-layer multi-hybrid expert group. Table 2: Table 2 lists the first activated sample feature data for each expert group, the semantic matching degree between each first activated sample feature data and the sample input data, and the first fusion weight for each expert group. Therefore, the first feature activation fusion result of the first layer multi-hybrid expert network is a fusion vector: The first feature activation fusion result is used as the input feature data of the second layer multi-hybrid expert model, and further feature fusion processing is performed in the second layer multi-hybrid expert model.

[0076] The following describes the analysis process in the second-level multi-hybrid expert model: In the second-layer multi-hybrid expert model, based on the first feature activation fusion result and the second fusion weight corresponding to each expert group in the second multi-hybrid expert model, the second activated sample feature data corresponding to each of the multiple expert groups are obtained. Here, a more advanced feature activation process is carried out on the basis of the first feature activation fusion result.

[0077] Specifically, the second fusion weights corresponding to each expert group in the second-layer multi-hybrid expert model are fused with the second activated sample feature data corresponding to each expert group to obtain the second feature activation fusion result of the second-layer multi-hybrid expert model. The specific calculation process is similar to that of the first-layer multi-hybrid expert model and will not be described here. Correspondingly, the second fusion weight G2 corresponding to each expert group in the second-layer multi-hybrid expert model is also obtained by back-calculation from the interaction behavior effective data prediction loss value obtained from the initial interaction behavior effective data prediction model, and is used to represent the importance of the second activated sample feature data of that expert group to the second feature activation fusion result.

[0078] Among them, the first fusion weight G1 and the second fusion weight G2 are learnable parameters. They are continuously adjusted by estimating the loss value. When the estimated loss value of the effective data of the interaction behavior is less than the preset threshold, the adjustment of the first fusion weight G1 and the second fusion weight G2 is stopped.

[0079] After obtaining the second feature activation fusion result, the effective data of the first predicted sample interaction behavior is obtained based on the fusion weight of the second feature activation fusion result and the first predicted data; the effective data of the second predicted sample interaction behavior is obtained based on the fusion weight of the second feature activation fusion result and the second predicted data. The effective data of the first predicted sample interaction behavior can be the estimated click-through rate data of sample users for the product description information of sample merchants, and the effective data of the second predicted sample interaction behavior can be the estimated order rate data of sample users for the products provided by sample merchants.

[0080] Step S103: Obtain the estimated effective data of the interaction behavior output by the initial effective data prediction model and the estimated loss value of the effective data of the sample interaction behavior.

[0081] This step is used to calculate the loss value between the estimated effective data of interactive behavior and the sample effective data of interactive behavior. It can be done by calculating the first estimated loss value between the estimated click-through rate data and the sample click-through rate data, calculating the second estimated loss value between the estimated order rate data and the sample order rate data, and obtaining the estimated loss value of the effective data of interactive behavior based on the first estimated loss value and the second estimated loss value.

[0082] After obtaining the estimated loss value of effective data of interaction behavior, it is used to calculate the adjustable parameters in the initial effective data of interaction behavior prediction model, together with the entropy regularization loss value obtained in the subsequent step S104.

[0083] Step S104: Based on the estimated effective data of the interaction behavior and the weight of the estimated effective data of the interaction behavior, obtain the entropy regularization loss value corresponding to the estimated effective data of the interaction behavior.

[0084] This step calculates the entropy regularization loss value corresponding to the effective data of the predicted interaction behavior. It analyzes the diversity of activated sample feature data for each expert group during the initial interaction behavior effective data prediction model's analysis of the input sample data. The purpose of this step is to make the model handle the types of activated sample feature data for each expert group more diverse during multiple training rounds, thereby improving the accuracy of parameter tuning for the model.

[0085] The positive entropy loss value is calculated as follows: Based on the weights of the estimated effective interactive behavior data, the current prediction entropy error value of the activated sample feature data is obtained; based on the estimated effective interactive behavior data, the perceptual weights corresponding to the sample input data are obtained; based on the perceptual weights and the current entropy error value, the entropy regularization loss value corresponding to the estimated effective interactive behavior data is obtained.

[0086] The current prediction entropy error value refers to the difference between the current prediction entropy value corresponding to the activated sample feature data and the target entropy value, which is obtained in the following way: Based on the weights of the estimated effective interactive behavior data corresponding to the estimated effective interactive behavior data, the activation probability of the activated sample feature data is obtained; based on the activation probability and the logarithm of the activation probability, the current predicted entropy value of the activated sample feature data is obtained; the target entropy value of the activated sample feature data is obtained; based on the current predicted entropy value and the target entropy value, the error value of the current predicted entropy value of the activated sample feature data is obtained.

[0087] The current predicted entropy value and the target entropy value are obtained through the following formulas (1) and (2), respectively: Formula (1) Formula (2) In formula (1), This represents the activation probability of each activated sample feature data when the model analyzes and processes the current sample input data. For example, the activation probabilities of the 5 activated sample feature data are g=[0.9,0.2,0.5,0.8,0.8]. Among them, the activation probabilities of the 1st, 4th, and 5th activated sample feature data are relatively high, indicating that the 1st, 4th, and 5th activated sample feature data are frequently activated and used during the multiple rounds of training of the model on other sample input data.

[0088] This is the logarithm of the activation probability of the activated sample feature data.

[0089] It represents the current prediction entropy value corresponding to the activated sample feature data in the current sample input data, which indicates the uniformity of the activation of the currently activated sample feature data.

[0090] The activated sample feature data in the current sample input data is the entropy value under uniform activation conditions.

[0091] The difference between the target entropy value and the current predicted entropy value is used to obtain the error value of the current predicted entropy value. ,if The larger the value, the smaller the current prediction entropy error, indicating a greater uniformity in the activation of sample features in the current input data. This means the activation frequency of the activated sample features is more evenly distributed across multiple training rounds. For example, during 10 training rounds, the probability of activating the first merchant feature in the merchant group is 4-6 times, which is relatively even, rather than 8-10 times. The former indicates a more even activation frequency for the first merchant feature during model analysis compared to the latter.

[0092] After obtaining the current prediction entropy error value, the entropy regularization loss value corresponding to the effective data of the predicted interaction behavior is obtained by combining the perceptual weights corresponding to the sample input data. This can be obtained by the following formula (3): Formula (3) in, This represents the entropy regularization loss value corresponding to the effective data of the predicted interaction behavior. The perceptual weights corresponding to the sample input data represent the importance of the sample input data in obtaining effective data for predicting interactive behavior. This indicates that the model's predicted interaction behavior is based on valid data. This indicates valid data representing the sample interaction behavior of the model; This indicates the relationship between the estimated loss value of the effective interaction data and the estimated loss value threshold. If the estimated loss value of the effective interaction data is greater than the estimated loss value threshold, then... If the estimated loss value of the effective data for the interaction behavior is less than the estimated loss value threshold, then... It is 0.

[0093] This means that formula (3) calculates the entropy regularization loss value for sample data where the estimated loss value of the effective data of the interaction behavior is greater than the threshold of the estimated loss value. This indicates that there is an obvious error in the model's prediction result and the model parameters need to be further adjusted.

[0094] By calculating the entropy regularization loss value, we can obtain the diversity of activated sample feature data when the model processes sample input data. The purpose is to encourage the model to activate a balance of sample feature data on sample data where the model prediction results are large.

[0095] Step S105: Obtain the sample real feature data in the sample input data. Based on the positive weight of the negative impact of the sample real feature data on the order fulfillment success result, obtain the group-level gradient gating weight of the activated sample feature data corresponding to the sample input data. The group-level gradient gating weight is the negative weight of the activated sample feature data on the order fulfillment success result. The order corresponding to the order fulfillment success result is an order generated based on the interaction behavior between the user and the merchant. The activated sample feature data is the sample feature data selected from the sample feature data corresponding to the sample input data to obtain the estimated effective data of the interaction behavior and the semantic alignment loss value.

[0096] This step is used in the model's reverse path to obtain the group-level gradient gating weights corresponding to multiple expert groups, respectively. Figure 3 Directional pathway section 303.

[0097] Real-valued feature data refers to the feature data in the sample input that affects the success of order fulfillment. Examples include delivery time data, merchant inventory levels, and delivery prices. This data influences the user's ability to successfully place an order for the merchant's food and receive the food.

[0098] Correspondingly, non-real-valued feature data in samples refers to feature data that has no impact on the order fulfillment result. For example, the merchant's identification information, which is a unique identifier provided by the system to the merchant, does not affect the success of order fulfillment.

[0099] The positive weight of the negative impact of the sample real feature data on the order fulfillment success result is: , which represents the positive weight of the sample real feature data that has a negative impact on the order fulfillment success result.

[0100] For example, a user likes a specific type of food from a merchant, but delivery to that item takes one hour. The user still orders the same food, and the meal arrives an hour later. In this scenario, the specific food item offered by the merchant positively impacts order fulfillment, while the one-hour delivery time negatively impacts it. The calculation here is... This refers to the positive weight of the negative impact of each sample's real-valued feature data on the order fulfillment success result.

[0101] After obtaining the positive weights of the negative impact of the sample real feature data on order fulfillment success, the group-level gradient gating weights are calculated by combining them with the anomaly scores corresponding to the sample real feature data. Specifically, the group-level gradient gating weights can be obtained as follows: Obtain the moving average and moving variance of the sample real feature data at the current time point; obtain the anomaly score corresponding to the sample real feature data based on the moving average and moving variance of the sample real feature data at the current time point; obtain the group-level gradient gating weight of the activated sample feature data corresponding to the sample input data based on the positive weight of the negative impact of the sample real feature data on the order fulfillment success result, including: obtaining the group-level gradient gating weight based on the positive weight of the negative impact of the sample real feature data on the order fulfillment success result and the anomaly score corresponding to the sample real feature data.

[0102] Please refer to formula (4): Formula (4) in, d represents the inverse weight of the negative impact of activated sample feature data in an expert group on the order fulfillment success outcome, where d represents an expert group. The value ranges from (0, 1), meaning it includes 1 but excludes 0. If the activated sample feature data of an expert group has a greater negative impact on order fulfillment success, then... The closer the value is to 0, the smaller the negative impact of the activated sample feature data of an expert group on order fulfillment success. The closer the value is to 1.

[0103] The sigmoid function is used to convert... The value is compressed to between 0 and 1, and in this application, it represents the probability value of the positive weight of the negative impact of the sample real feature data on the successful order fulfillment.

[0104] This represents the anomaly score corresponding to the real number features of the sample.

[0105] It represents a constant.

[0106] Among them, the anomaly scores corresponding to the real features of the samples It can be obtained through the following formula (5): Formula (5) in, This represents the moving average of the sample's real-valued features at the current point in time. This represents the moving variance of the sample's real-valued feature data at the current time point; This indicates the current value of the real-valued feature data of the currently input sample; It represents a constant.

[0107] For example, if the real-valued feature data of the sample being analyzed is delivery time, then... Specific data indicating the current delivery time, for example, 1 hour; This represents the average delivery time at the current point in time, for example, 0.5 hours.

[0108] The current moving average of the sample real-valued feature data at the current time point is obtained in the following way: Obtain the actual observed value of the sample real feature data at the current time point and the previous estimated moving average value corresponding to the previous time point; based on the previous estimated moving average value and the first average weight corresponding to the previous estimated moving average value, and the actual observed value and the second average weight corresponding to the actual observed value, obtain the current estimated moving average value of the sample real feature data at the current time point, where the sum of the first average weight and the second average weight is a first preset weight value.

[0109] Please refer to formula (6): Formula (6) in, This represents the current moving average of the sample's real-valued feature data at the current point in time. This represents the previous estimated moving average of the sample's real-valued feature data at the previous time point corresponding to the current time point. This indicates the weight of the first average value corresponding to the previous estimated moving average; This represents the actual observed value of the sample's real-valued feature data at the current time point; This represents the weight of the second average value corresponding to the actual observed value; The sum of the first average weight and the second average weight is the first preset weight value. For example, it can be 1 or any other value. It can be set according to the specific requirements of the scheme, and there is no restriction here.

[0110] The current moving variance value of the sample real feature data at the current time point is obtained in the following way: Obtain the absolute value of the moving difference between the actual observed value and the current estimated moving average of the sample real feature data at the current time point; obtain the previous moving variance value corresponding to the previous time point of the sample real feature data at the current time point; obtain the current moving variance value of the sample real feature data at the current time point based on the previous moving variance value and the first variance value weight corresponding to the previous moving variance value, and the absolute value of the moving difference at the current time point and the second variance value weight corresponding to the absolute value of the moving difference at the current time point, wherein the current moving variance value is used to represent the degree of fluctuation of the sample real feature data at the current time point, and the sum of the first variance value weight and the second variance value weight is a second preset weight value.

[0111] Please refer to formula (7): Formula (7) in, This represents the current moving variance of the sample's real feature data at the current time point; This represents the moving variance value of the sample real feature data at the previous time point corresponding to the current time point; This indicates the weight of the first variance value corresponding to the previous moving variance value; This represents the absolute value of the moving difference between the actual observed value and the current estimated moving average of the sample's real feature data at the current time point. This represents the weight of the second squared difference corresponding to the absolute value of the shift difference at the current time point.

[0112] The sum of the weight of the first variance value and the weight of the second variance value is the second preset weight value. For example, it can be 1, or it can be other values. There are no restrictions here.

[0113] The above describes the data process involved in obtaining the group-level gradient gating weights corresponding to the expert group in this step. By obtaining the group-level gradient gating weights, the parameter update gradient obtained by back-calculating the effective data loss value and entropy regularization loss value through interactive behavior is controlled to reduce the parameter update gradient of the corresponding expert group. This will be analyzed in step S106.

[0114] Step S106: Adjust the initial interactive behavior effective data prediction model based on the effective data prediction loss value of the interactive behavior, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights to obtain the trained interactive behavior effective data prediction model.

[0115] This step is used to adjust the initial interaction behavior effective data prediction model based on the interaction behavior effective data prediction loss value, entropy regularization loss value, semantic alignment loss value, and group-level gradient gating weights obtained in the above steps.

[0116] Specifically as follows: Based on the estimated loss value of the effective interaction data, the entropy regularization loss value, and the semantic alignment loss value, the original gradient data corresponding to the correlation parameters that need to be adjusted in the initial effective interaction data estimation model is obtained. The activated sample feature data corresponding to the original gradient data is then obtained, where the correlation parameters to be adjusted are the correlation parameters corresponding to the category to which the activated sample feature data belongs. Based on the group-level gradient gating weights corresponding to the activated sample feature data and the original gradient data, the updated gradient data of the activated sample feature data corresponding to the original gradient data is obtained. Based on the updated gradient data, the updated weights of the correlation parameters to be adjusted are obtained. The correlation parameters are then adjusted based on the updated weights.

[0117] Please refer to formula (8): Formula (8) in, This represents the total loss value. Based on the total loss value, the parameters that need to be adjusted during model training are determined. This represents the estimated loss value based on valid data of the interaction behavior. This represents the entropy regularization loss value; This represents the weight corresponding to the entropy regularization loss value; This represents the semantic alignment loss value; This represents the weight corresponding to the semantic alignment loss value.

[0118] According to formula (8), the total loss value of the model is obtained here. Based on the total loss value, the correlation parameters that need to be adjusted in the model are obtained by back-reasoning through the chain rule. The correlation parameters that need to be adjusted are the correlation parameters corresponding to the category to which the activated sample feature data belongs, such as the first fusion weight or the second fusion weight corresponding to the above multiple expert groups respectively.

[0119] Furthermore, based on the total loss value, the first total loss value for the effective data of the first estimated interaction behavior (estimated click-through rate data) and the second total loss value for the effective data of the second estimated behavior (estimated order rate data) are derived.

[0120] The correlation parameters related to the estimated click-through rate (CTR) data are obtained by back-calculating the first total loss value. These correlation parameters correspond to the category to which the activated sample feature data related to the estimated CTR data belongs.

[0121] For example, the activated sample feature data associated with the estimated click-through rate data is delivery time data. Delivery time data belongs to the sample real feature data of the constraint group, and the correlation parameter that needs to be adjusted can be the first fusion weight or the second fusion weight of the constraint group.

[0122] By taking the derivative of the first total loss value, the original gradient data corresponding to the required adjustment of the correlation parameters is obtained.

[0123] The updated gradient data with the required adjustment parameters is obtained by combining the original gradient data with the group-level gradient gating weights corresponding to the category of the activated sample feature data to which the associated parameter belongs.

[0124] Please refer to formula (9): Formula (9) in, This represents the updated gradient data; This represents the original gradient data corresponding to the association parameter of the category to which the activated sample feature data belongs; This represents the inverse weight of the negative impact of the category to which the activated sample feature data belongs on the order fulfillment success outcome; y=1 indicates that the order was successfully fulfilled; y=0 indicates that the order has failed to be fulfilled.

[0125] This approach analyzes the total loss between the model's predicted results and the actual results for successful order fulfillment based on sample input data. The required adjustment of correlation parameters is then determined based on this total loss. For sample feature data that positively impacts successful order fulfillment, the original gradient of the correlation parameters for the corresponding categories is increased; conversely, for sample feature data that negatively impacts successful order fulfillment, the original gradient is decreased.

[0126] Based on the above example, since delivery time (1 hour) is a feature data point that abnormally impacts order fulfillment success, the gradient update required for the associated parameters (e.g., the first fusion weight) of the category (expert group) corresponding to the delivery time data in the sample input data should be reduced in the model parameter update process to obtain a reduced gradient update amount. If the delivery time data belongs to a constraint group, the gradient data of the associated parameters corresponding to the constraint group should also be reduced, as should the gradient data of the associated parameters corresponding to the shared group. This can reduce the abnormal impact of noisy feature data on the model parameter update process and avoid overfitting of the model parameters. For example, if the original gradient data is 5 and the group-level gradient gating weight is 0.2, then the updated gradient data will be 0.2. 5 = 1.

[0127] Furthermore, in step S102, obtaining the semantic alignment loss value between the identifier information of the sample merchant and the textual semantic description information of the sample merchant output by the initial interaction behavior effective data prediction model includes: The candidate identifier information of the sample merchants, the basic description semantic vector of the sample merchants, and the context constraint difference semantic vector are provided to the semantic alignment model, and the semantic alignment model performs the following steps: Obtain the candidate identifier vector corresponding to the candidate identifier information of the sample merchant; obtain the first semantic similarity between the candidate identifier vector of the sample merchant and the basic description semantic vector; obtain the second semantic similarity between the candidate identifier vector of the sample merchant and the context constraint difference semantic vector; based on the first semantic similarity and the second semantic similarity, perform semantic vector transfer processing on the candidate identifier vector of the sample merchant with the basic description semantic vector and the context constraint difference semantic vector, respectively.

[0128] Both the basic description semantic vector and the context constraint difference semantic vector of the sample merchants are obtained from a frozen large model. The candidate identifier vectors of the sample merchants are continuously adjusted by comparing the first semantic similarity between the candidate identifier vector and the basic description semantic vector, and the second semantic similarity between the candidate identifier vector and the context constraint difference semantic vector. This process continues until the first semantic similarity is less than a first similarity threshold and the second semantic similarity is less than a second semantic similarity threshold. At this point, the basic description semantic vector and the context constraint difference semantic vector of the sample merchants are determined to be matched with the candidate identifier information of the sample merchants. Therefore, semantic vector transfer processing is performed, specifically fusing the identifier vector, basic description semantic vector, and context constraint difference semantic vector corresponding to the sample merchant. This allows the model to obtain the basic description information and differentiated description information of the sample merchant based on its identifier information.

[0129] For cold-start merchants, i.e., merchants who are joining the service platform for the first time from online to offline, these merchants have no historical evaluation records or user records on the service platform. By providing these merchants with rich descriptive information through the method of this solution, when a user makes a request message on the service platform, the system can obtain the descriptive information of the sample merchant and the effective data of the estimated behavior between the user and the merchant through this model, thereby increasing the probability of recommending the merchant to the user.

[0130] The interactive behavior effective data prediction model training method provided in this application determines the parameters that the model needs to adjust by using the interactive behavior effective data prediction loss value, and determines the update magnitude of the parameters that the model needs to adjust by combining group-level gradient gating weights. Entropy regularization loss value is used to determine the diversity of activated analysis data in the sample input data. Semantic alignment loss value is used to improve the semantic transfer process between the sample merchant's identification information and textual semantic description information, thereby improving the model's understanding of the service information that the sample merchant can provide to the user. Therefore, the method provided in this application can improve the accuracy of the trained model in analyzing the effective data of interactive behavior between the user and the sample merchant, thereby improving the user experience.

[0131] Second Embodiment Based on the first embodiment, the second embodiment of this application provides a training device for an effective data prediction model of interactive behavior, such as... Figure 4 As shown, the effective data prediction model for interaction behavior is used to predict effective data of user-merchant interaction behavior. The device includes: The acquisition unit 401 is used to acquire sample data for training the effective data prediction model of interactive behavior. The sample data includes sample input data and sample output data. The sample input data includes the identification information of the sample merchant and the text semantic description information of the sample merchant. The sample output data is the sample effective data of interactive behavior expected to be obtained based on the sample input data to represent the effectiveness of interactive behavior. The model analysis unit 402 is used to provide the sample input data to the initial interaction behavior effective data prediction model, obtain the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model, and obtain the semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial interaction behavior effective data prediction model. The interaction behavior effective data prediction loss value acquisition unit 403 is used to obtain the interaction behavior effective data prediction loss value between the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model and the sample interaction behavior effective data. The entropy regularization loss value acquisition unit 404 is used to obtain the entropy regularization loss value corresponding to the estimated interaction behavior effective data based on the estimated interaction behavior effective data and the estimated interaction behavior effective data weight corresponding to the estimated interaction behavior effective data. The group-level gradient gating weight acquisition unit 405 is used to acquire sample real feature data in the sample input data, and obtain the group-level gradient gating weight of the activated sample feature data corresponding to the sample input data based on the positive weight of the negative impact of the sample real feature data on the order fulfillment success result. The group-level gradient gating weight is the negative weight of the activated sample feature data on the negative impact of the order fulfillment success result. The order corresponding to the order fulfillment success result is an order generated based on the interaction behavior between the user and the merchant. The activated sample feature data is the sample feature data selected from the sample feature data corresponding to the sample input data to obtain the estimated effective data of the interaction behavior and the semantic alignment loss value. Training unit 406 is used to adjust the initial interactive behavior effective data prediction model based on the interactive behavior effective data prediction loss value, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights, so as to obtain the trained interactive behavior effective data prediction model.

[0132] Third Embodiment Based on the first embodiment, the third embodiment of this application provides a method for obtaining effective interactive behavior data. This method is applied to electronic devices, such as servers, desktop computers, laptops, smart TVs, VR devices, in-vehicle devices, wearable devices, mobile phones, tablets, smartwatches, and other electronic devices with data processing functions.

[0133] Please refer to Figure 5 The method for obtaining effective interactive behavior data provided in the third embodiment of this application includes the following steps S501 to S503.

[0134] The method provided in the third embodiment of this application is applicable to a first application scenario, such as detecting a user's triggering operation on a target application and displaying recommended service information and a list of candidate merchants providing the service information on the target application's page. The service information and the list of candidate merchants can be obtained in advance in the following manner, or obtained in real time after detecting the user's triggering operation on the target application: Based on the user's historical log information in the target application, service information that can be recommended to the user and candidate merchant information that can provide such service information are determined. According to the prediction model provided in the first embodiment, valid data on the interaction behavior between the user and each candidate merchant is obtained, such as estimated order rate data or estimated click-through rate data. Based on the valid data on the interaction behavior between the user and each candidate merchant, the information of multiple candidate merchants is sorted according to a preset sorting rule to obtain a list of candidate merchant information.

[0135] The method provided in the third embodiment of this application is also applicable to a second application scenario. For example, when a user's operation to obtain target service information is detected in a target application, a list of candidate merchants that can provide the target service information is displayed on the target application's page. Specifically, based on the user's operation to obtain the target service information, candidate merchant information that can provide the target service information is obtained. Then, in conjunction with the first application scenario type, the prediction model provided in the first embodiment is used to obtain valid data on the interaction behavior between the user and each candidate merchant. Based on the valid data on the interaction behavior between the user and each candidate merchant, the information of multiple candidate merchants is sorted according to a preset sorting rule to obtain a list of candidate merchant information.

[0136] The above are merely examples illustrating the methods provided in this application. The methods provided in the third embodiment of this application can also be applied to other application scenarios, and are not limited here.

[0137] Step S501: Obtain the service information required by the user and the information of candidate merchants who can provide the service information; The service information and candidate merchant information obtained in this step can be obtained through the methods described above, or through other methods, which are not limited here.

[0138] Step S502: The service information required by the user and the candidate merchant information are used as input data and provided to the trained interaction behavior effective data prediction model to obtain the effective data of interaction behavior between the user and the candidate merchant output by the interaction behavior effective data prediction model. The effective data prediction model for interactive behavior is obtained by training using the training method described in the first embodiment.

[0139] The effective data prediction model of interactive behavior is used to analyze the effective data of the interaction behavior between the user and multiple candidate merchants, such as the predicted click-through rate data or the predicted order rate data.

[0140] Step S503: Based on the valid data of the interaction behavior between the user and the candidate merchants, sort the information of multiple candidate merchants to obtain the sorting information corresponding to the multiple candidate merchant information respectively.

[0141] This step sorts the candidate merchants from highest to lowest value based on the valid interaction data corresponding to each merchant, obtaining a ranking for each candidate merchant. Based on the ranking information, a list of candidate merchants is generated and provided to the user, who can then select a target merchant from the list to provide the desired service.

[0142] Fourth embodiment Based on the third embodiment, the fourth embodiment of this application provides a device for obtaining effective data on interactive behavior, such as... Figure 6 As shown, the device 600 for obtaining valid interactive behavior data includes: The first information acquisition unit 601 is used to acquire the service information required by the user and the information of candidate merchants that can provide the service information. The effective data acquisition unit 602 is used to provide the service information required by the user and the candidate merchant information as input data to the trained effective data prediction model of interactive behavior, and to obtain the effective data of interactive behavior between the user and the candidate merchant output by the effective data prediction model of interactive behavior. The effective data prediction model of interactive behavior is trained by the training method described in the first embodiment. The sorting unit 603 is used to sort multiple candidate merchant information based on the valid data of the interaction behavior between the user and the candidate merchant, and obtain sorting information corresponding to the multiple candidate merchant information respectively.

[0143] Fifth Embodiment Corresponding to the method of the first embodiment of this application, the fifth embodiment of this application also provides an electronic device. This electronic device includes: a processor; and a memory for storing a computer program. After the electronic device is powered on and the processor runs the computer program, it executes the method of the first embodiment. Figure 7 As shown, Figure 7 This is a schematic diagram of an electronic device provided in the fifth embodiment of this application. The electronic device specifically includes: at least one processor 701, at least one communication interface 702, at least one memory 703, and at least one communication bus 704. Optionally, the communication interface 702 can be an interface for a communication module, such as the interface for a GSM module. The processor 701 may be a CPU, an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The memory 703 may include high-speed RAM and may also include non-volatile memory, such as at least one disk storage device. The memory 703 stores a program, and the processor 701 calls the program stored in the memory 703 to execute the above-described method embodiments of the present invention.

[0144] Sixth Embodiment The sixth embodiment of this application also provides a computer storage medium storing a computer program, which is executed by a processor to perform the above-described method embodiments of the present invention.

[0145] The above-described device embodiments, electronic device embodiments, and storage medium embodiments correspond to the above-described method embodiments; please refer to the method embodiments for details. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application. In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory. Memory may include non-permanent storage in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media. 1. Computer-readable media includes permanent and non-permanent, removable and non-removable media. Information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change 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), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media do not include non-transitory computer-readable media, such as modulated data signals and carrier waves. 2. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. It should be noted that the embodiments of this application may involve the use of user data. In practical applications, user-specific personal data may be used in the scheme described herein within the scope permitted by applicable laws and regulations, provided that it complies with the applicable laws and regulations of the country (e.g., with the user's explicit consent, with the user being properly notified, etc.).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 entry points are provided for users to choose to authorize or refuse.

Claims

1. A method for training an effective data prediction model for interactive behavior, characterized in that, The effective data prediction model for interaction behavior is used to predict effective data of user-merchant interaction behavior, and the method includes: Obtain sample data for training an effective data prediction model of interactive behavior. The sample data includes sample input data and sample output data. The sample input data includes the identification information of the sample merchants and the textual semantic description information of the sample merchants. The sample output data is the sample effective data of interactive behavior expected to be obtained based on the sample input data to represent the effectiveness of interactive behavior. The sample input data is provided to the initial interaction behavior effective data prediction model to obtain the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model, and the semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial interaction behavior effective data prediction model is obtained. Obtain the interaction behavior effective data prediction loss value between the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model and the sample interaction behavior effective data; Based on the estimated effective data of interactive behavior and the weight of the estimated effective data of interactive behavior, the entropy regularization loss value corresponding to the estimated effective data of interactive behavior is obtained. The sample real feature data in the sample input data is obtained. Based on the positive weight of the negative impact of the sample real feature data on the order fulfillment success result, the group-level gradient gating weight of the activated sample feature data corresponding to the sample input data is obtained. The group-level gradient gating weight is the negative weight of the activated sample feature data on the order fulfillment success result. The order corresponding to the order fulfillment success result is the order generated based on the interaction behavior between the user and the merchant. The activated sample feature data is the sample feature data selected from the sample feature data corresponding to the sample input data to obtain the estimated effective data of the interaction behavior and the semantic alignment loss value. Based on the effective data prediction loss value of the interaction behavior, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights, the initial effective data prediction model of the interaction behavior is adjusted to obtain the trained effective data prediction model of the interaction behavior.

2. The method according to claim 1, characterized in that, The sample input data includes sample merchant feature data, which includes the merchant's identification information and textual semantic description information; the sample merchant feature data is obtained in the following way: Obtain the basic description information of the sample merchants; The basic description information of the sample merchants is provided to the frozen big model to obtain the basic description semantic vector of the sample merchants and the candidate merchant feature vector library. The candidate merchant feature vector library includes the basic description semantic vectors corresponding to multiple merchants generated by the frozen big model. Based on the context information and constraints corresponding to the service categories that the sample merchants can provide to users, a preset number of candidate merchant basic description semantic vectors associated with the sample merchants are selected from the candidate vector library. Based on the basic description semantic vector of the sample merchant and the basic description semantic vector of the candidate merchant, obtain the context constraint difference semantic vector between the sample merchant and the candidate merchant; The basic description semantic vector of the sample merchant and the context constraint difference semantic vector are used as the feature data of the sample merchant.

3. The method according to claim 1, characterized in that, The sample input data includes sample input data from multiple categories; The step of providing the sample input data to the initial interaction behavior effective data prediction model and obtaining the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model includes: The sample input data of the multiple categories are provided to the initial interaction behavior effective data prediction model, and the following steps are performed through the initial interaction behavior effective data prediction model: Obtain the sample feature data corresponding to the sample input data for each category, and use it as the sample feature data for each category; Obtain the first activated sample feature data and the first inactive sample feature data in the sample feature data of each category; The feature data of the first activated samples and the feature data of the first non-activated samples of all categories are fused to obtain the first feature activation fusion result; Based on the first feature activation fusion result and the weights corresponding to the sample feature data of each category, the second activated sample feature data and the second unactivated sample feature data in the sample feature data of each category are obtained; The feature data of the second activated samples and the feature data of the second non-activated samples of all categories are fused to obtain the second feature activation fusion result; Based on the results of the second feature activation fusion, the effective data for the predicted interaction behavior is obtained.

4. The method according to claim 1, characterized in that, The step of adjusting the initial interaction behavior effective data prediction model based on the estimated loss value of the interaction behavior effective data, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights includes: Based on the effective data prediction loss value of the interaction behavior, the entropy regularization loss value, and the semantic alignment loss value, the original gradient data corresponding to the correlation parameters that need to be adjusted in the initial effective data prediction model of the interaction behavior is obtained, and the activated sample feature data corresponding to the original gradient data is obtained. The correlation parameters that need to be adjusted are the correlation parameters corresponding to the category to which the activated sample feature data belongs. Based on the group-level gradient gating weights corresponding to the activated sample feature data corresponding to the original gradient data and the original gradient data, the updated gradient data of the activated sample feature data corresponding to the original gradient data is obtained. Based on the updated gradient data, obtain the update weights of the related parameters that need to be adjusted; The associated parameters are adjusted based on the updated weights.

5. The method according to claim 1, characterized in that, The step of obtaining the entropy regularization loss value corresponding to the estimated interaction behavior effective data based on the estimated interaction behavior effective data and the corresponding weights of the estimated interaction behavior effective data includes: Based on the weights of the estimated effective interactive behavior data corresponding to the estimated effective interactive behavior data, the current prediction entropy error value of the activated sample feature data is obtained. The perceptual weights corresponding to the sample input data are obtained based on the effective data of the predicted interactive behavior. Based on the perception weights and the current entropy error value, the entropy regularization loss value corresponding to the effective data of the predicted interaction behavior is obtained.

6. A training device for an effective data prediction model of interactive behavior, characterized in that, The effective data prediction model for interactive behavior is used to predict effective data of user-merchant interaction behavior. The device includes: The acquisition unit is used to acquire sample data for training an effective data prediction model of interactive behavior. The sample data includes sample input data and sample output data. The sample input data includes the identification information of the sample merchant and the textual semantic description information of the sample merchant. The sample output data is the sample effective data of interactive behavior expected to be obtained based on the sample input data to represent the effectiveness of interactive behavior. The model analysis unit is used to provide the sample input data to the initial interaction behavior effective data prediction model, obtain the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model, and obtain the semantic alignment loss value between the identifier information of the sample merchant and the text semantic description information of the sample merchant output by the initial interaction behavior effective data prediction model. The interaction behavior effective data prediction loss value acquisition unit is used to obtain the interaction behavior effective data prediction loss value between the predicted interaction behavior effective data output by the initial interaction behavior effective data prediction model and the sample interaction behavior effective data. The entropy regularization loss value acquisition unit is used to obtain the entropy regularization loss value corresponding to the estimated interaction behavior effective data based on the estimated interaction behavior effective data and the estimated interaction behavior effective data weight corresponding to the estimated interaction behavior effective data. A group-level gradient gating weight acquisition unit is used to acquire sample real-valued feature data in the sample input data, and obtain the group-level gradient gating weight of the activated sample feature data corresponding to the sample input data based on the positive weight of the negative impact of the sample real-valued feature data on the order fulfillment success result. The group-level gradient gating weight is the negative weight of the activated sample feature data on the negative impact of the order fulfillment success result. The order corresponding to the order fulfillment success result is an order generated based on the interaction behavior between the user and the merchant. The activated sample feature data is the sample feature data selected from the sample feature data corresponding to the sample input data to obtain the effective data of the estimated interaction behavior and the semantic alignment loss value. The training unit is used to adjust the initial interactive behavior effective data prediction model based on the interactive behavior effective data prediction loss value, the entropy regularization loss value, the semantic alignment loss value, and the group-level gradient gating weights, so as to obtain the trained interactive behavior effective data prediction model.

7. A method for obtaining effective data on interactive behavior, characterized in that, The method includes: Obtain the service information required by the user and information on candidate merchants who can provide the service information; The user's required service information and the candidate merchant information are used as input data and provided to the trained interaction behavior effective data prediction model to obtain the effective data of interaction behavior between the user and the candidate merchant output by the interaction behavior effective data prediction model. The interaction behavior effective data prediction model is trained by the training method described in any one of claims 1 to 5. Based on the valid data of the interaction behavior between the user and the candidate merchants, the information of multiple candidate merchants is sorted to obtain the sorting information corresponding to each of the multiple candidate merchant information.

8. A device for obtaining effective data on interactive behavior, characterized in that, The device includes: The information acquisition unit is used to acquire the service information required by the user and the information of candidate merchants that can provide the service information. The effective data acquisition unit for interactive behavior is used to provide the service information required by the user and the candidate merchant information as input data to the trained effective data prediction model for interactive behavior, and to obtain the effective data of interactive behavior between the user and the candidate merchant output by the effective data prediction model for interactive behavior, wherein the effective data prediction model for interactive behavior is trained by the training method described in any one of claims 1 to 5. The sorting unit is used to sort multiple candidate merchant information based on the valid data of the interaction behavior between the user and the candidate merchant, and obtain sorting information corresponding to the multiple candidate merchant information respectively.

9. An electronic device, characterized in that, include: Memory and processor; The memory is used to store one or more computer instructions; The processor is used to execute one or more computer instructions to implement the method as described in any one of claims 1-5 and 7.

10. A computer-readable storage medium storing one or more computer instructions thereon, characterized in that, The instruction is executed by the processor to implement the method as described in any one of claims 1-5, 7.

Images