Object generation model training method and device, computer equipment, readable storage medium and program product

Abstract

The invention relates to an object generation model training method and device, computer equipment, a computer readable storage medium and a computer program product. The method comprises the following steps: acquiring training data; performing weighting processing based on the multi-source behavior data, each behavior type weight and the time attenuation weight to obtain a static interest score; coding the training data to obtain user portrait features, group portrait features and behavior sequence features, and obtaining an attention result; processing is carried out through the attention result and a feedforward neural network, and a dynamic prediction interest score is generated; and training the to-be-trained object generation model through the sample interest score, the static interest score, the dynamic prediction interest score and the joint loss function to obtain the object generation model which effectively fuses the multi-source data and has good performance.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a training method, apparatus, computer device, computer-readable storage medium, and computer program product for an object generation model. Background Technology

[0002] With the rapid development of artificial intelligence technology, users are increasingly browsing and using different types of insurance products via the internet. As internet technology further develops in the insurance field, user insurance behavior data is becoming increasingly multi-source and frequent.

[0003] In related technologies, user interest tags for various products are generally generated solely from static user data. This relies on behavioral data types and machine learning models to identify user intent, resulting in low accuracy of the generated user tags. Summary of the Invention

[0004] Therefore, it is necessary to provide a training method, apparatus, computer device, computer-readable storage medium, and computer program product that can integrate multi-source data to generate an object generation model with accurate labels, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for training an object generation model, comprising:

[0006] Acquire training data, which includes user profile data, user group profile data, multi-source behavioral data, and sample interest scores. The multi-source behavioral data includes user behavior data in various interactive components.

[0007] The static interest score is obtained by weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight; and the training data is encoded to obtain user profile features, group profile features, and behavioral sequence features.

[0008] Attention is applied to the user profile features, group profile features, and behavioral sequence features to obtain attention results; the attention results are then processed with a feedforward neural network to generate a dynamic predicted interest score.

[0009] The object generation model is trained by using the sample interest score, the static interest score, the dynamic predicted interest score, and the joint loss function.

[0010] In some embodiments, the method further includes:

[0011] Acquire initial behavioral data, which includes user behavior data in various interactive components;

[0012] The initial behavioral data is cleaned and standardized to obtain multi-source behavioral data, which includes behavioral type, behavioral time, and related behavioral data.

[0013] In some embodiments, the step of weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight to obtain a static interest score includes:

[0014] For each behavior type, a weighted average is performed based on the statistical data of the behavior type, the weight of the behavior type, and the time decay weight corresponding to the behavior type to obtain a behavior type score;

[0015] The static interest score of the user for each target is obtained by superimposing the scores of each behavior type.

[0016] In some embodiments, the attention mechanism processing based on the user profile features, group profile features, and behavioral sequence features to obtain an attention result; and the processing of the attention result and a feedforward neural network to generate a dynamically predicted interest score, including:

[0017] The user profile features and the group profile features are concatenated to obtain a concatenated vector, and the concatenated vector is projected to determine the projection result as the query vector; and a key vector and a value vector are obtained based on the behavior sequence features.

[0018] Based on the query vector, key vector, and value vector, an attention mechanism is applied to obtain the attention weight vector;

[0019] The attention weight vector is processed by a first feedforward neural network to obtain the attention result;

[0020] Based on the second feedforward neural network and the normalization layer, the attention results and multi-source behavioral data are processed to generate dynamic interest scores for users for each target.

[0021] In some embodiments, the step of training the object generation model to be trained using the sample interest score, the static interest score, the dynamically predicted interest score, and the joint loss function to obtain a trained object generation model includes:

[0022] The sample interest score, the static interest score, and the dynamically predicted interest score are processed by a joint loss function to obtain loss data;

[0023] If the preset training completion conditions are not met, the current object generation model is trained based on the loss data to obtain an updated object generation model. Based on the updated object generation model, the step of obtaining training data is repeated until the preset training completion conditions are met to obtain a trained object generation model.

[0024] In some embodiments, the joint loss function includes one or more of a regression loss function, a ranking consistency loss function, a group constraint loss function, and a regularization loss function; the loss data is calculated using one or more of the following:

[0025] The difference data between the sample interest score and the static interest score and the dynamic predicted interest score is calculated using the mean squared error function, and the difference data is determined to be the loss data corresponding to the regression loss function.

[0026] The sample interest score, the static interest score, and the dynamically predicted interest score are processed using the ranking consistency loss function to obtain the loss data corresponding to the ranking consistency loss function.

[0027] Calculate the difference data between the static interest score, the dynamic predicted interest score, and the interest score of the user's user group, and determine the difference data as the loss data of the group constraint loss function;

[0028] The model parameters of the current object generation model are processed based on the preset regularization mode and regularization parameters to obtain the loss data corresponding to the regularization loss function.

[0029] Secondly, this application also provides a training apparatus for an object generation model, comprising:

[0030] The acquisition module is used to acquire training data, which includes user profile data, user group profile data, multi-source behavior data, and sample interest scores. The multi-source behavior data includes user behavior data in various interactive components.

[0031] The static score determination module is used to perform weighted processing based on the multi-source behavioral data, the weights of each behavioral type, and the time decay weight to obtain a static interest score; and to encode the training data to obtain user profile features, group profile features, and behavioral sequence features.

[0032] The dynamic score determination module is used to perform attention mechanism processing based on the user profile features, group profile features and behavioral sequence features to obtain attention results; and to generate dynamic predicted interest scores by processing the attention results and feedforward neural networks.

[0033] The training module is used to train the object generation model to be trained using the sample interest scores, the static interest scores, the dynamic predicted interest scores, and the joint loss function, so as to obtain a trained object generation model.

[0034] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0035] Acquire training data, which includes user profile data, user group profile data, multi-source behavioral data, and sample interest scores. The multi-source behavioral data includes user behavior data in various interactive components.

[0036] The static interest score is obtained by weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight; and the training data is encoded to obtain user profile features, group profile features, and behavioral sequence features.

[0037] Attention is applied to the user profile features, group profile features, and behavioral sequence features to obtain attention results; the attention results are then processed with a feedforward neural network to generate a dynamic predicted interest score.

[0038] The object generation model is trained by using the sample interest score, the static interest score, the dynamic predicted interest score, and the joint loss function.

[0039] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0040] Acquire training data, which includes user profile data, user group profile data, multi-source behavioral data, and sample interest scores. The multi-source behavioral data includes user behavior data in various interactive components.

[0041] The static interest score is obtained by weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight; and the training data is encoded to obtain user profile features, group profile features, and behavioral sequence features.

[0042] Attention is applied to the user profile features, group profile features, and behavioral sequence features to obtain attention results; the attention results are then processed with a feedforward neural network to generate a dynamic predicted interest score.

[0043] The object generation model is trained by using the sample interest score, the static interest score, the dynamic predicted interest score, and the joint loss function.

[0044] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0045] Acquire training data, which includes user profile data, user group profile data, multi-source behavioral data, and sample interest scores. The multi-source behavioral data includes user behavior data in various interactive components.

[0046] The static interest score is obtained by weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight; and the training data is encoded to obtain user profile features, group profile features, and behavioral sequence features.

[0047] Attention is applied to the user profile features, group profile features, and behavioral sequence features to obtain attention results; the attention results are then processed with a feedforward neural network to generate a dynamic predicted interest score.

[0048] The object generation model is trained by using the sample interest score, the static interest score, the dynamic predicted interest score, and the joint loss function.

[0049] The above-mentioned object generation model training method, apparatus, computer device, computer-readable storage medium, and computer program product, wherein the method includes: acquiring training data, including user profile data, user group profile data, multi-source behavior data, and sample interest scores, wherein the multi-source behavior data includes user behavior data in various interactive components; performing weighted processing based on the multi-source behavior data, the weights of each behavior type, and the time decay weight to obtain static interest scores; encoding the training data to obtain user profile features, group profile features, and behavior sequence features; performing attention mechanism processing based on the user profile features, group profile features, and behavior sequence features to obtain attention results; processing the attention results and a feedforward neural network to generate dynamic predicted interest scores; and training the object generation model to be trained using the sample interest scores, static interest scores, dynamic predicted interest scores, and a joint loss function to obtain a trained object generation model.

[0050] This method effectively integrates multi-source data, enabling accurate identification of user intent by fully utilizing multi-dimensional data. It constructs dual-track static and dynamic interest evaluation scores through multi-source behavioral data, ensuring the comprehensiveness of interest scores. Furthermore, it combines multi-level modeling and multi-head attention mechanisms to model the object generation model, improving its performance and ensuring real-time response to changes in user interests. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart illustrating the training method of an object generation model in one embodiment;

[0053] Figure 2 This is a schematic diagram illustrating the attention mechanism calculation in one embodiment;

[0054] Figure 3 This is a structural block diagram of a training device for an object generation model in one embodiment;

[0055] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0058] In one embodiment, such as Figure 1As shown, a training method for an object generation model is provided. This embodiment illustrates the application of this method to a server. It is understood that this method can also be applied to terminals, and to systems including terminals and servers, and is implemented through interaction between the terminal and the server. The terminals can be, but are not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. The server can be a standalone server or a server cluster consisting of multiple servers. In this embodiment, the training method for the object generation model includes the following steps:

[0059] Step 102: Obtain training data.

[0060] The training data includes user profile data, user group profile data, multi-source behavioral data, and sample interest scores. The multi-source behavioral data includes user behavior data in various interactive components. User profile data refers to data describing the individual characteristics of a user, such as geographical location, spending power, usage frequency, active time periods, gender, etc. Usage frequency refers to how often a user uses a target component, and active time periods refer to the time periods during which a user prefers to use the target component. User group profile data refers to the characteristics of the group to which a user belongs, such as age group, gender, etc. Multi-source behavioral data can be a sequence of user interaction behaviors on multiple target components. This sequence includes multiple behaviors and the time of occurrence of each behavior. Each target component can be a component that displays data to be recommended, such as an application, a mini-program in social software, or a payment platform, etc. The interaction behavior sequence includes multiple interaction behavior data, which can be the same type of interaction behavior or different types of interaction behavior. This interaction behavior can include exposure, click, browsing, sharing, conversion, etc. This interaction behavior data can include behavior attributes and values, timestamps, and page types. Behavior attributes refer to the type of behavior, values ​​refer to the duration or frequency of the behavior, such as the duration of browsing behavior, etc.; timestamps are the time when the behavior occurred or the duration of the behavior, and page types can be the page where the behavior occurred, such as the application's splash screen, main page, or search page, etc. The sample interest scores include users' historical interest scores for each target to be recommended.

[0061] Specifically, after determining the object generation model to be trained, the server can obtain user profile data, user group profile data, real-time multi-source behavioral data, and user sample interest scores for the target to be recommended, corresponding to the object generation model. For example, the server can collect the above information from multiple data source components (target components), which may include insurance applications, mini-programs within social applications, payment platforms, and user profile systems, etc. Data from the user profile system can be automatically synchronized to the server, and each target component can report data to the server through a data tracking SDK (Software Development Kit).

[0062] Step 104 involves weighting the multi-source behavioral data, the weights of each behavioral type, and the time decay weight to obtain a static interest score. The training data is then encoded to obtain user profile features, group profile features, and behavioral sequence features.

[0063] Among these, the behavior type weight represents the degree to which a particular behavior type influences a user's interest level; for example, different types of behavior have different impacts on the calculation of a user's interest score. For instance, the behavior type weight for the "collect" behavior related to the target to be recommended is greater than that for the "browse" behavior. The time decay weight represents the degree to which the time of the interaction influences a user's interest level; for example, recently occurring behaviors have a higher time decay weight. The static interest score represents the user's static baseline interest score for the target to be recommended.

[0064] Specifically, the server can identify various user behaviors related to the target to be recommended from multi-source behavioral data, and determine the statistical data for each behavior (such as the number of clicks for click behavior, the number of searches for search behavior, etc.), as well as the behavior type weight and the time decay weight corresponding to that behavior. The server can calculate the time decay data by using a preset decay coefficient, time decay weight, and preset exponential function. Thus, for each of the multiple user behaviors related to the target to be recommended, the server can determine the initial static score corresponding to that behavior based on the product of the behavior type weight, behavior statistical data, and time decay data. The server can also calculate the sum of the initial static scores of each behavior and determine the sum as the user's static interest score for the target to be recommended.

[0065] For example, the server can use a preset encoding algorithm to encode user profile data, user group profile data, and multi-source behavior data in the training data to obtain user profile features, group profile features, and behavior sequence features.

[0066] Optionally, the server can perform feature encoding on user profile data and user group profile data to obtain user profile features and group profile features. These user profile features and group profile features can be normalized and discretized feature vectors. Correspondingly, the server can transform the behavior sequences in multi-source behavior data, converting the structured triplet information into a structure of behavior type, timestamp, and associated content. This associated content can represent behaviors related to the behavior; for example, the associated content for a click behavior can be browsing behavior, representing the mutual association between the click behavior and the browsing behavior.

[0067] Step 106: Based on user profile features, group profile features, and behavioral sequence features, an attention mechanism is applied to obtain the attention result. The attention result is then processed using a feedforward neural network to generate a dynamically predicted interest score.

[0068] The object generation model may include a data preprocessing module, an attention module, and a decoder, wherein the attention module includes at least a feedforward neural network.

[0069] Specifically, in the attention module of the object generation model, the server can process user profile features, group profile features, and behavioral sequence features using an attention mechanism to generate attention results. The attention results are then processed by a feedforward neural network in the object generation model to obtain the output of the feedforward neural network. The output of the feedforward neural network is then decoded by a decoder to obtain the model output. Based on the model output, the user's dynamic predicted interest score for the target to be recommended is determined.

[0070] Step 108: Train the object generation model to be trained using sample interest scores, static interest scores, dynamic predicted interest scores, and a joint loss function to obtain a trained object generation model.

[0071] Specifically, the server can process one or more of the following—sample interest scores, static interest scores, and dynamically predicted interest scores—using a joint loss function to generate a joint loss value for the object generation model. If the training completion condition is not met, the parameters of the object generation model are updated using this joint loss value, resulting in an updated model. Based on the updated model, the process of acquiring training data is repeated until the preset training completion condition is met, resulting in a trained object generation model.

[0072] In one example, when the server determines that the current number of training iterations has met the preset training iteration threshold, the server can determine that the preset training completion condition has been met. In another example, when the server determines that the current calculation of the joint loss value has met the preset convergence condition, the server can determine that the preset training completion condition has been met. The preset convergence condition may be that the joint loss value has remained unchanged, or that the joint loss value has reached the minimum loss value threshold, etc. This embodiment does not limit the specific value of the minimum loss value threshold, and those skilled in the art can determine it according to the actual application scenario.

[0073] In the training method of the above object generation model, training data is acquired, including user profile data, user group profile data, multi-source behavior data, and sample interest scores. The multi-source behavior data includes user behavior data in various interactive components. A weighted average is performed based on the multi-source behavior data, the weights of each behavior type, and the time decay weight to obtain a static interest score. The training data is then encoded to obtain user profile features, user group profile features, and behavior sequence features. An attention mechanism is applied to the user profile features, user group profile features, and behavior sequence features to obtain an attention result. The attention result and a feedforward neural network are then processed to generate a dynamic predicted interest score. Finally, the object generation model to be trained is trained using the sample interest score, static interest score, dynamic predicted interest score, and a joint loss function to obtain a trained object generation model.

[0074] This method effectively integrates multi-source data, enabling accurate identification of user intent by fully utilizing multi-dimensional data. It constructs dual-track static and dynamic interest evaluation scores through multi-source behavioral data, ensuring the comprehensiveness of interest scores. Furthermore, it combines multi-level modeling and multi-head attention mechanisms to model the object generation model, improving its performance and ensuring real-time response to changes in user interests.

[0075] In some embodiments, the method further includes:

[0076] Initial behavioral data is acquired, and then cleaned and standardized to obtain multi-source behavioral data.

[0077] Initial behavioral data includes user behavior data across various interactive components. Multi-source behavioral data includes behavior type, behavior time, and related behavioral data.

[0078] Specifically, the user behavior-related data collected directly by the server from various target components can be unstructured log data, i.e., initial behavior data. The server performs anomaly filtering on the initial behavior data, removing any abnormal behaviors. Alternatively, the server can perform structuring processing, converting the unstructured log data into structured triple information. This triple information includes behavior type, page type, and timestamp, where the page type represents the page on which the behavior occurred. Correspondingly, the server can also perform timestamp alignment processing on the initial behavior data, for example, aligning the timestamps of each target component (interactive component) and each terminal with the server's time as the base time. Based on one or more of the above preprocessing operations, the initial behavior data is converted into multi-source behavior data.

[0079] Optionally, the server can identify an abnormal behavior as the number of operations within a preset time period exceeding a preset number threshold, or as the behavior with a dwell time less than a preset time threshold; the preset time period can be one minute, the preset number threshold can be 50 times, and the preset time threshold can be 1 second.

[0080] In this embodiment, data preprocessing is performed to ensure the uniformity of data format, which facilitates subsequent model understanding.

[0081] In one embodiment, the specific processing steps for "weighting based on multi-source behavioral data, weights of each behavioral type, and time decay weights to obtain a static interest score" include:

[0082] For each behavior type, a weighted average is calculated based on the behavior type's statistical data, weight, and corresponding time decay weight to obtain a behavior type score. These scores are then combined to calculate a user's static interest score for each target.

[0083] Among them, the statistical data of behavior types can be data obtained after statistics are collected on a certain type of behavior. For example, it can be the browsing time and number of browsing behaviors, or the number of clicks for click behaviors, the number of searches for search behaviors, and so on.

[0084] Specifically, the server can identify various user behaviors related to the target to be recommended from multi-source behavioral data, and determine the statistical data for each behavior type, as well as the corresponding behavior type weight and time decay weight. The server can calculate the time decay data using a preset decay coefficient, time decay weight, and preset exponential function. Thus, for each of the user's multiple behaviors related to the target to be recommended, the server can determine the initial static score corresponding to that behavior based on the product of the behavior type weight, behavior statistical data, and time decay data. The server can then calculate the sum of the initial static scores for each behavior, determining this sum as the user's static interest score for the target to be recommended. The target can be the target to be recommended, such as different types of insurance products, etc.

[0085] For example, a user's static interest scores for each target. It can be calculated using the following formula:

[0086]

[0087] Where i represents the i-th action performed by the user towards the target, behavior type i represents the behavior statistics of the i-th behavior type, weight i can represent the behavior type weight of the i-th behavior, λ can be a preset decay coefficient, ti can be the time decay weight corresponding to the time of behavior occurrence, and the preset exponential function can be an exponential function with base e. It can characterize time decay data.

[0088] In this embodiment, by combining time decay weight and behavior type weight to calculate the static baseline interest score, the long-term stable level of user interest in the target to be recommended can be accurately calculated.

[0089] In one embodiment, the specific processing steps of "processing attention based on user profile features, group profile features, and behavioral sequence features to obtain attention results, and processing the attention results and feedforward neural networks to generate dynamically predicted interest scores" include:

[0090] User profile features and group profile features are concatenated to obtain a concatenated vector. This concatenated vector is then projected, and the resulting projection is used as the query vector. Key and value vectors are obtained based on behavioral sequence features. An attention mechanism is applied to the query vector, key vector, and value vector to obtain an attention weight vector. This attention weight vector is then processed by a first feedforward neural network to obtain the attention result. Finally, based on a second feedforward neural network and a normalization layer, the attention result and multi-source behavioral data are processed to generate dynamic interest scores for users regarding various objectives.

[0091] The dynamic interest score can be a dynamic dual-track interest score. The object generation model can include a data preprocessing module, an attention mechanism module, and a decoder. The attention mechanism module can include an attention mechanism unit and a feedforward neural network unit. The attention mechanism unit includes a first normalization layer, a multi-head attention mechanism layer, and a first feedforward neural network; the feedforward neural network unit includes a second normalization layer and a second feedforward neural network.

[0092] Specifically, the output of the data preprocessing module can be the input data of the attention mechanism module, i.e., it includes at least multi-source behavioral data. In the attention mechanism module, the server can normalize the user profile features, group profile features, and behavioral sequence features through the first normalization layer to obtain normalized user profile features, group profile features, and behavioral sequence features. Then, in the multi-head attention mechanism layer, the user profile features and group profile features are concatenated / connected to obtain a concatenated vector, and the concatenated vector is projected to obtain a projected vector, which is determined as the query vector (query, Q). Correspondingly, the server can determine the behavioral sequence features as key vector K (key) and value vector V (value). The product between the query vector and the key vector is calculated, and the product is normalized and processed by Softmax (activation function) to obtain the activation result. The product between the activation result and the value vector is calculated, and the product is determined as the attention weight vector. The attention weight vector is then superimposed on the input data of the attention mechanism module through the first feedforward neural network to obtain the superposition result, which is determined as the attention result. The server can process the attention results in the second normalization layer to obtain a normalized result. This normalized result is then processed by a second feedforward neural network to obtain a first vector. The product between the attention result and a preset value is calculated, and the sum of this product and the first vector is determined. This summed result is identified as the output of the attention mechanism module. The decoder then decodes this summed result to obtain the dynamic interest score predicted by the object generation model for each target to be recommended. The preset value can be determined based on the total number of targets to be recommended; for example, the total number can be set as the preset value.

[0093] The attention weight vector Attention(Q,K,V) can be calculated using the following formula:

[0094] ;

[0095] Among them, K T d represents the transpose of K; k It is the dimension of the K vector (scaling to avoid gradient vanishing).

[0096] Optionally, the attention mechanism module may include N attention modules connected end to end. Each attention module includes an attention mechanism unit and a feedforward neural network unit. The specific execution process of each unit is consistent with the process described in the above embodiment, and will not be repeated here. The server can input the output of the Nth attention module into the decoder to obtain the dynamic interest score of the user for each target to be recommended, as predicted by the object generation model.

[0097] Optionally, the server can determine long-term static interest scores s based on the user's static interest scores for each target. l and short-term static interest scores s s The server can also process the attention results and feedforward neural networks to generate long-term dynamic predicted interest scores d. l-dynamic and short-term dynamic prediction of interest score d s-dynamic In this way, the server can calculate the preset coefficient b1 and the short-term dynamic predicted interest score d. s-dynamic The product between them is calculated, and this product is then compared with the long-term static interest score s. l The sum of the values ​​between the two values ​​determines the user's short-term preference score d for the target to be recommended. s Accordingly, the server can calculate the preset coefficient b2 and the long-term dynamic predicted interest score d. l-dynamic The product between them is calculated, and this product is then compared with the short-term static interest score s. s The sum of the values ​​between the two values ​​determines the user's long-term preference score d for the target to be recommended. l Specifically, it can be calculated using the following formula:

[0098] ;

[0099] .

[0100] In this embodiment, the attention mechanism is used to identify the temporal dependencies of the combination patterns of individual user behaviors, which can identify the association between behavioral context and dynamic interests. Furthermore, by adjusting the behavior weights, refined modeling of different groups can be achieved, and the correlation between different groups and behaviors can be improved. Accurate evaluation of user interests can be achieved through dynamic scores, and the K vector can be scaled by the dimension of the statistical key vector to avoid gradient vanishing.

[0101] In one embodiment, the specific processing steps of "training the object generation model to be trained using sample interest scores, static interest scores, dynamically predicted interest scores, and a joint loss function to obtain a trained object generation model" include:

[0102] The sample interest score, static interest score, and dynamically predicted interest score are processed by a joint loss function to obtain loss data. If the preset training completion conditions are not met, the current object generation model is trained based on the loss data to obtain an updated object generation model. Based on the updated object generation model, the step of obtaining training data is repeated until the preset training completion conditions are met, resulting in a trained object generation model.

[0103] Specifically, the server can process one or more of the following—sample interest scores, static interest scores, and dynamically predicted interest scores—using a joint loss function to generate a joint loss value for the object generation model. If the training completion condition is not met, the parameters of the object generation model are updated using this joint loss value, resulting in an updated model. Based on the updated model, the process of acquiring training data is repeated until the preset training completion condition is met, resulting in a trained object generation model.

[0104] In one example, when the server determines that the current number of training iterations has met the preset training iteration threshold, the server can determine that the preset training completion condition has been met. In another example, when the server determines that the current calculation of the joint loss value has met the preset convergence condition, the server can determine that the preset training completion condition has been met. The preset convergence condition may be that the joint loss value has remained unchanged, or that the joint loss value has reached the minimum loss value threshold, etc. This embodiment does not limit the specific value of the minimum loss value threshold, and those skilled in the art can determine it according to the actual application scenario.

[0105] In this embodiment, the loss value of the object generation model is calculated by using a joint loss function, which can ensure the comprehensiveness of the loss value calculation, improve the model training efficiency, and further improve the accuracy of the model output results.

[0106] In one embodiment, the joint loss function includes one or more of the following: regression loss function, ranking consistency loss function, group constraint loss function, and regularization loss function. Loss data is calculated using one or more of the following:

[0107] The mean squared error function is used to calculate the difference between the sample interest score and the static interest score and the dynamically predicted interest score, and the difference data is determined as the loss data corresponding to the regression loss function.

[0108] By processing the sample interest scores, static interest scores, and dynamically predicted interest scores using the ranking consistency loss function, the loss data corresponding to the ranking consistency loss function is obtained.

[0109] Calculate the difference data between the static interest score, the dynamically predicted interest score, and the interest score of the user's user group, and determine the difference data as the loss data of the group constraint loss function.

[0110] The model parameters of the current object generation model are processed based on the preset regularization mode and regularization parameters to obtain the loss data corresponding to the regularization loss function.

[0111] Among them, the regression loss function is the main loss function, which is used to align with the true labels numerically; the ranking consistency loss function is used to align the predicted distribution and the label distribution, and enhance the normality of the relative order of importance; the group constraint loss function can make the preferences of users in the same group more similar; and the regularization loss can optimize the training process, avoid the weights or interest scores from being too concentrated, and avoid the parameters from overfitting.

[0112] Specifically, the server can obtain sample interest scores from the training data. These sample interest scores represent the user's true interest scores for each target to be recommended, determined based on historical data. The server can calculate the difference data between the sample interest scores and the static interest scores, and between the sample interest scores and the dynamically predicted interest scores, respectively, using the mean squared error function (MSE). The sum of these difference data is then used as the loss data corresponding to the regression loss function.

[0113] The server can determine the first distribution of users' interest in each target to be recommended based on sample interest scores. The server can also determine the second distribution of users' interest in each target to be recommended based on static interest scores and dynamically predicted interest scores, as predicted by the object generation model. The first and second distributions are calculated using the cross-entropy loss function to obtain the loss data corresponding to the ranking consistency loss function.

[0114] The server can obtain the degree of interest of the user's group towards each target to be recommended, and process the static interest score, dynamic predicted interest score, and the degree of interest of the user's group towards each target to be recommended through the group constraint loss function to obtain the loss data of the group constraint loss function.

[0115] The server can determine that the regularization mode can include a first regularization mode or a second regularization mode, such as L2 regularization mode; the model parameters of the model generated for the current object are processed by the regularization parameters corresponding to the L2 regularization mode to obtain the loss data corresponding to the regularization loss function.

[0116] Based on this, the server can perform weighted processing using one or more loss data points and the weights corresponding to each loss data point to obtain a weighted result, and determine the weighted result as the joint loss value corresponding to the joint loss function.

[0117] In this embodiment, the loss value of the object generation model is calculated by using a joint loss function, which can ensure the comprehensiveness of the loss value calculation, improve the model training efficiency, and further improve the accuracy of the model output results.

[0118] The following describes in detail the training process of the above-mentioned object generation model, using a specific embodiment. This model can be applied to the dynamic behavior tag generation server of an insurance recommendation system. It integrates multi-source user behavior data and generates dynamic interest tags through a complete process: "data acquisition → preprocessing → multi-level feature modeling → attention mechanism processing → model training → tag generation." The insurance recommendation system may include a data acquisition module, a data preprocessing module, a multi-level feature modeling module, an attention mechanism module, and a model training optimization module. This insurance recommendation system acquires user features and behavioral features, and performs dual-track dynamic interest score modeling and static baseline interest score modeling based on these features. The user features include user profiles and behavioral sequences. User profiles are static tags, and behavioral sequences include exposure, clicks, browsing, sharing, and conversions. The dual-track dynamic interest score modeling includes multi-dimensional complete pattern modeling and refined group feature modeling. Multi-dimensional complete pattern modeling uses an attention mechanism to effectively identify combined behavioral patterns and temporal dependencies, and capture personally important and abrupt dynamic patterns. Fine-grained modeling of group characteristics, based on the different weight distributions of different groups, such as the greater importance of conversion behavior of high-purchasing-power groups to category preference, and the greater importance of search and click behavior of potential product users to category preference, guides the modeling of interest scores; the process of dual-track dynamic interest score modeling and static baseline interest score modeling may include:

[0119] Static baseline interest scores are modeled based on user and behavioral characteristics to obtain static interest scores, which may include short-term preference scores (interest scores) S. S and long-term preference S l Dynamic modeling based on user and behavioral characteristics can yield predicted short-term dynamic interest scores d. S-dynamic and long-term dynamic interest score d l-dynamic By weighting and fusing the above data, we can obtain a comprehensive understanding of users' short-term and long-term preferences for the target to be recommended. In this way, we can determine users' preference tags and push dynamic tags to the recommendation system via API, supporting millisecond-level latency.

[0120] In the data acquisition module, the server can collect data from sources such as insurance apps, WeChat mini-programs, third-party payment platforms, and customer relationship management (CRM) systems. It collects multi-source data in real time, including behavioral sequence data such as user actions on apps / mini-programs (including behavioral attributes and values, timestamps, page types, etc.). It can also collect user profile data, such as demographic attributes, geographical distribution, spending power, health information, policy amount, usage frequency, and active time periods. The data acquisition module can achieve data collection through the following methods: each terminal (app / mini-program / payment platform) reports data to the server via a data tracking SDK, and the customer relationship management system synchronizes data to the server.

[0121] In the server-side data preprocessing module, the raw data is cleaned and standardized to obtain the preparation data for modeling. This includes one or more of the following: anomaly filtering, standardization, timestamp alignment, and feature engineering. Anomaly filtering refers to removing robot behavior (rule: single IP > 50 clicks per minute) and accidental touch operations (identification dwell time < 1 second). Standardization refers to converting unstructured logs into structured triples (behavior type, page type, timestamp). Timestamp alignment refers to unifying the timestamps of data from multiple terminals based on the insurance APP server time (error ≤ 100ms). Feature engineering can be used to convert behavior sequences into a "behavior type + occurrence time + associated content" structure, and user profiles into normalized / discrete embedding vectors, for example, converting "monthly income > 20,000" into a vector [0.9, 0.1].

[0122] In the multi-level feature modeling module, static baseline interest scores and dynamic dual-track interest scores can be generated, taking into account both long-term stable interests and short-term dynamic changes. The static baseline interest score is obtained based on behavioral depth statistics (such as click count) × behavior type weight × time decay weight. For example, the time decay weight indicates that behaviors that have occurred recently will have a higher weight, such as behaviors in the last 7 days, reflecting the user's long-term stable interests. The static baseline interest score S 静态 It can be calculated using the formula:

[0123] ;

[0124] The dynamic dual-track interest score includes individual multidimensional complete pattern modeling and refined group feature modeling. Individual multidimensional complete pattern modeling uses an attention mechanism to identify the combination patterns and temporal dependencies of individual user behaviors. For example, it can identify the temporal correlation between various behaviors such as the personalized pattern of "searching for life insurance → clicking on critical illness insurance". Refined group feature modeling adjusts the behavioral weights for different groups (such as high-purchasing-power people and potential product users) (e.g., the conversion behavior weight of high-purchasing-power people is higher) to capture macroscopic isomorphic patterns.

[0125] like Figure 2 The diagram shown illustrates the structure of the attention mechanism module, which includes multiple normalization layers and multiple feedforward neural networks, including a multi-head attention mechanism layer, etc. The specific application process of this attention mechanism module has been described in detail in the above embodiments and will not be repeated here. This attention mechanism module can focus on calculating the relevance of Query(Q), Key(K), and Value(V), capturing behavioral context and dynamic interests through a multi-head attention mechanism, wherein:

[0126] The query vector Q is obtained by concatenating the user profile embedding and the group profile embedding, and then projecting the result onto an MLP (Multilayer Perceptron). The key vector K and value vector are obtained by normalizing the behavior sequence embedding. This step enhances the relevance of "high-purchasing-power individuals → payment behavior" and captures personalized interests. The attention weights Attention(Q,K,V) can be calculated using the following formula:

[0127] ;

[0128] Optionally, the object generation model is a multi-head attention neural network and a multi-loss constraint training model, which can ensure the model's generalization ability. The model structure includes multi-layer repeating modules (including feedforward neural network, layer normalization, and multi-head attention). The query, key, and value can also be divided into h heads (e.g., 8 heads), and attention can be computed in parallel. For example, the attention weights can be calculated using the following formula:

[0129] ;

[0130] The training methods for the object generation model can support full training (initial model training), incremental training (fine-tuning with new data), and online training (real-time model updates); the regression loss in the loss function can be replaced with Huber loss, and the ranking consistency loss can be replaced with KL divergence loss; the number of model layers in the object generation model: the number of multi-head attention layers N can be adjusted according to the data scale (e.g., increasing from 6 layers to 12 layers); the constraint loss of the object generation model can include one or more of the following:

[0131] Regression loss: Main loss, measured by mean squared error (MSE) to determine the numerical difference between the model's output interest score and the true label;

[0132] Ranking Consistency Loss: Aligns the prediction distribution with the label distribution using cross-entropy loss to enhance the relative order of interest importance;

[0133] Group constraint loss: encourages users in the same group to have similar preferences, thus improving the ability to model isomorphic patterns in groups;

[0134] Regularization loss: L2 regularization avoids overfitting of weights.

[0135] Optionally, the server can convert the interest scores output by the object generation model into dynamic interest tags for each target to be recommended by the user, and push the dynamic tags to the recommendation system via API, supporting millisecond-level latency.

[0136] The following example illustrates the training process of the object generation model described above. For instance, it could be the generation process of short-term intention tags for user A regarding the target "medical insurance intention":

[0137] S1, Data Collection: User A browses the "Million Medical Insurance" page on an insurance APP (stays for 120 seconds), completes the premium payment in a WeChat mini program, and the user profile shows that he / she is a "high purchasing power group" (monthly income > X).

[0138] S2, Preprocessing: Clean the data and standardize it to "Browse-Million Medical Insurance-202X-XX-XX 14:30:00" and "Payment-Million Medical Insurance-202X-XX-XX 14:35:00", aligning the timestamps.

[0139] S3, Feature Modeling: Static baseline scores are based on "browsing + payment" statistics; individual modeling captures the personalized pattern of "high-purchasing-power group → payment behavior"; group modeling enhances the weight of conversion behavior of "high-purchasing-power group".

[0140] S4, Attention Mechanism: The Q vector represents "high purchasing power + group profile", and the K / V vector represents "browsing + payment" behavior embedding. Attention score increases the weight of "payment" behavior.

[0141] S5, Model Training: Through multi-loss constraint optimization, the model outputs a "short-term medical insurance intention" score of 0.85.

[0142] S6, Tag Generation: Weighted fusion of short-term preferences, output "Short-term medical insurance intention (confidence level 0.85)", and push it to the recommendation system.

[0143] This embodiment provides a training method for an object generation model. It utilizes a multi-head attention mechanism, where multiple attention heads learn dependencies from different perspectives in parallel, and then merge the results to enhance the model's expressive power, thereby capturing richer and more multi-faceted dependencies in the input sequence. It can reflect stable user interests based on long-term user behavior statistics (e.g., click count) × behavior type weight × time decay weight. It can capture short-term changes in user interests through dynamic dual-track modeling (individual + group) and an attention mechanism. The correlation between behavior and user can be calculated using QKV vectors in the attention mechanism. Regression loss (e.g., mean squared error MSE) measures the numerical difference between the model's output interest score and the true label. Ranking consistency loss can also measure the consistency between the predicted interest distribution and the true label distribution; cross-entropy loss is commonly used.

[0144] This embodiment provides a training method for an object generation model that can integrate multi-source cross-platform insurance behavior data, construct a multi-level dynamic interest model through a static baseline and dual-track dynamic scores, capture behavioral context using a multi-head attention mechanism, and optimize model training with multiple loss constraints. It can make full use of labeled data, improve the robustness of sparse data, identify dynamic interests, and achieve real-time (millisecond-level updates), accuracy (capturing personalized patterns), and dynamic adaptability (responding to changes in interests) of user interest labels.

[0145] This embodiment provides a training method for an object generation model that can fully utilize multi-dimensional data: it integrates user profiles, behavior sequences, and static data to piece together complete user intent; it can improve the robustness of sparse data: through group feature modeling and attention mechanisms, it can infer deep intents from sparse data; it can capture behavioral associations: the attention mechanism captures behavioral context (such as the difference between "search → click" and "click → search"); it can identify dynamic interests: dual-track dynamic scores and incremental training respond to changes in user interests in real time; and it can also model isomorphic patterns in groups: group constraint loss improves the consistency of preferences within the same group.

[0146] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0147] Based on the same inventive concept, this application also provides a training apparatus for an object generation model to implement the above-described training method for the object generation model. The solution provided by this apparatus is similar to the implementation described in the above-described method. Therefore, the specific limitations of one or more object generation model training apparatus embodiments provided below can be found in the limitations of the object generation model training method described above, and will not be repeated here.

[0148] In one exemplary embodiment, such as Figure 3 As shown, a training device 300 for an object generation model is provided, comprising:

[0149] The acquisition module 302 is used to acquire training data, which includes user profile data, user group profile data, multi-source behavior data and sample interest scores. The multi-source behavior data includes user behavior data in various interactive components.

[0150] The static score determination module 304 is used to perform weighted processing based on the multi-source behavioral data, the weights of each behavioral type, and the time decay weight to obtain a static interest score; and to encode the training data to obtain user profile features, group profile features, and behavioral sequence features.

[0151] The dynamic score determination module 306 is used to perform attention mechanism processing based on the user profile features, group profile features and behavioral sequence features to obtain attention results; and to generate dynamic predicted interest scores by processing the attention results and the feedforward neural network.

[0152] The training module 308 is used to train the object generation model to be trained using the sample interest score, the static interest score, the dynamic predicted interest score, and the joint loss function, so as to obtain a trained object generation model.

[0153] In some embodiments, the apparatus further includes:

[0154] The initial data acquisition module is used to acquire initial behavior data, which includes user behavior data in various interactive components;

[0155] The standardization processing module is used to clean and standardize the initial behavioral data to obtain multi-source behavioral data, wherein the behavioral data in the multi-source behavioral data includes behavioral type, behavioral time, and related behavioral data.

[0156] In some embodiments, the static score determination module is specifically used for:

[0157] For each behavior type, a weighted average is performed based on the statistical data of the behavior type, the weight of the behavior type, and the time decay weight corresponding to the behavior type to obtain a behavior type score;

[0158] The static interest score of the user for each target is obtained by superimposing the scores of each behavior type.

[0159] In some embodiments, the dynamic score determination module is specifically used for:

[0160] The user profile features and the group profile features are concatenated to obtain a concatenated vector, and the concatenated vector is projected to determine the projection result as the query vector; and a key vector and a value vector are obtained based on the behavior sequence features.

[0161] Based on the query vector, key vector, and value vector, an attention mechanism is applied to obtain the attention weight vector;

[0162] The attention weight vector is processed by a first feedforward neural network to obtain the attention result;

[0163] Based on the second feedforward neural network and the normalization layer, the attention results and multi-source behavioral data are processed to generate dynamic interest scores for users for each target.

[0164] In some embodiments, the training module is specifically used for:

[0165] The sample interest score, the static interest score, and the dynamically predicted interest score are processed by a joint loss function to obtain loss data;

[0166] If the preset training completion conditions are not met, the current object generation model is trained based on the loss data to obtain an updated object generation model. Based on the updated object generation model, the step of obtaining training data is repeated until the preset training completion conditions are met to obtain a trained object generation model.

[0167] In some embodiments, the joint loss function includes one or more of a regression loss function, a ranking consistency loss function, a group constraint loss function, and a regularization loss function; the loss data is calculated using one or more of the following:

[0168] The difference data between the sample interest score and the static interest score and the dynamic predicted interest score is calculated using the mean squared error function, and the difference data is determined to be the loss data corresponding to the regression loss function.

[0169] The sample interest score, the static interest score, and the dynamically predicted interest score are processed using the ranking consistency loss function to obtain the loss data corresponding to the ranking consistency loss function.

[0170] Calculate the difference data between the static interest score, the dynamic predicted interest score, and the interest score of the user's user group, and determine the difference data as the loss data of the group constraint loss function;

[0171] The model parameters of the current object generation model are processed based on the preset regularization mode and regularization parameters to obtain the loss data corresponding to the regularization loss function.

[0172] The modules in the training device for the aforementioned object generation model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0173] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a training method for an object generation model. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0174] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0175] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0176] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0177] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0178] 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, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0179] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0180] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0181] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method of training an object generation model, the method comprising: The method comprises: obtaining training data, the training data comprising user portrait data, user group portrait data, multi-source behavior data and sample interest scores, the multi-source behavior data comprising user behavior data in each interactive component; based on the multi-source behavior data, each behavior type weight and the time decay weight, performing weighted processing to obtain a static interest score; and encoding the training data to obtain user portrait features, group portrait features and behavior sequence features; based on the user portrait features, group portrait features and behavior sequence features, performing attention mechanism processing to obtain an attention result; and through the attention result and a feedforward neural network, performing processing to generate a dynamic predicted interest score; through the sample interest scores, the static interest scores, the dynamic predicted interest scores and a joint loss function, performing training processing on an object generation model to be trained to obtain a trained object generation model.

2. The method of claim 1, wherein, The method further comprises: obtaining initial behavior data, the initial behavior data comprising user behavior data in each interactive component; performing cleaning processing and standardization processing on the initial behavior data to obtain multi-source behavior data, the behavior data in the multi-source behavior data comprising behavior types, behavior times and associated behavior data.

3. The method of claim 1, wherein, The weighted processing based on the multi-source behavior data, each behavior type weight and the time decay weight to obtain a static interest score comprises: for each behavior type, based on the statistical data of the behavior type, the behavior type weight and the time decay weight corresponding to the behavior type, performing weighted processing to obtain a behavior type score; based on each behavior type score, performing superposition calculation to obtain a static interest score of the user for each target.

4. The method of claim 1, wherein, The attention mechanism processing based on the user portrait features, group portrait features and behavior sequence features to obtain an attention result comprises: splicing the user portrait features and the group portrait features to obtain a spliced vector, performing projection processing on the spliced vector to determine a projection result as a query vector; and based on the behavior sequence features, obtaining a key vector and a value vector; based on the query vector, key vector and value vector, performing attention mechanism processing to obtain an attention weight vector; through a first feedforward neural network, performing processing on the attention weight vector to obtain an attention result; based on a second feedforward neural network and a normalization layer, performing processing on the attention result and multi-source behavior data to generate a dynamic interest score of the user for each target. The training processing on an object generation model to be trained through the sample interest scores, the static interest scores, the dynamic predicted interest scores and a joint loss function to obtain a trained object generation model comprises:

5. The method of claim 1, wherein, through a joint loss function, performing processing on the sample interest scores, the static interest scores and the dynamic predicted interest scores to obtain loss data; ​ In the case where the preset training completion condition is not met, the current object generation model is trained based on the loss data to obtain an updated object generation model, and the step of obtaining training data is re-executed based on the updated object generation model until the preset training completion condition is met, and a trained object generation model is obtained.

6. The method of claim 5, wherein, The joint loss function comprises one or more of a regression loss function, a ranking consistency loss function, a group constraint loss function, and a regularization loss function; and the loss data is calculated by one or more of the following: The difference data between the sample interest score, the static interest score, and the dynamic predicted interest score is calculated by a mean square error function, and the difference data is determined as the loss data corresponding to the regression loss function; The sample interest score, the static interest score, and the dynamic predicted interest score are processed by a ranking consistency loss function to obtain loss data corresponding to the ranking consistency loss function; The difference data between the static interest score, the dynamic predicted interest score, and the interest score of the user group to which the user belongs is calculated, and the difference data is determined as the loss data of the group constraint loss function; The model parameters of the current object generation model are processed based on a preset regularization mode and a regularization parameter to obtain loss data corresponding to the regularization loss function.

7. An apparatus for training an object generation model, the apparatus comprising: The device comprises: An acquisition module configured to acquire training data, the training data comprising user portrait data, user group portrait data, multi-source behavior data, and a sample interest score, the multi-source behavior data comprising user behavior data in each interactive component; A static score determination module configured to perform weighted processing based on the multi-source behavior data, each behavior type weight, and a time decay weight to obtain a static interest score, and to encode the training data to obtain user portrait features, group portrait features, and behavior sequence features; A dynamic score determination module configured to perform attention mechanism processing based on the user portrait features, group portrait features, and behavior sequence features to obtain an attention result, and to perform processing on the attention result and a feedforward neural network to generate a dynamic predicted interest score; A training module configured to perform training processing on a to-be-trained object generation model based on the sample interest score, the static interest score, the dynamic predicted interest score, and a joint loss function to obtain a trained object generation model. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images