Interaction behavior prediction model training method and device, equipment, medium and product

Abstract

Embodiments of the present application disclose an interactive behavior estimation model training method, device, equipment, medium and product. The method comprises: obtaining first interactive behavior feature information and second interactive behavior feature information corresponding to an interactive behavior sample set; then determining feature association mode information between heterogeneous features based on the second interactive behavior feature information; and performing completion processing on missing features in the first interactive behavior feature information based on the feature association mode information to obtain interactive behavior completion feature information; and further performing model training on an initial interactive behavior estimation model based on the interactive behavior completion feature information and an interactive behavior label corresponding to the interactive behavior sample set to obtain a target interactive behavior estimation model.

Description

Technical Field

[0001] This application relates to the field of model training technology, and in particular to a method, apparatus, device, medium and product for training an interactive behavior prediction model. Background Technology

[0002] In applications such as product recommendation, content recommendation, and advertising, it is often necessary to predict the probability of a user performing interactive behaviors such as clicking on a recommended target based on multiple user characteristics. However, due to factors such as sparse user behavior, incomplete data collection, and untimely data synchronization across systems, the feature information input to the model may be missing to varying degrees. When the proportion of missing features is high, the interactive behavior prediction models in related technologies struggle to adapt to the differences in data structure and variation patterns of different types of features, resulting in low accuracy of missing feature processing results. This further affects the accuracy, stability, and generalization ability of the interactive behavior prediction results. Summary of the Invention

[0003] This application provides a method, apparatus, device, medium, and product for training an interactive behavior prediction model, which can improve the generalization ability and prediction accuracy of the interactive behavior prediction model. The above technical solution is as follows: In a first aspect, embodiments of this application provide a method for training an interactive behavior prediction model, including: Obtain first and second interactive behavior feature information corresponding to the interactive behavior sample set. The interactive behavior sample set includes multiple heterogeneous features. The first interactive behavior feature information corresponds to the original value space, and the second interactive behavior feature information corresponds to the embedding space. Based on the second interaction behavior feature information, determine the feature association pattern information between heterogeneous features; Based on the feature association pattern information, the missing features in the first interaction behavior feature information are completed to obtain the interaction behavior completion feature information. Based on the feature information of interactive behavior completion and the interactive behavior labels corresponding to the interactive behavior sample set, the initial interactive behavior prediction model is trained to obtain the target interactive behavior prediction model.

[0004] In one possible implementation, obtaining the first interaction behavior feature information and the second interaction behavior feature information corresponding to the interaction behavior sample set includes: Obtain the interaction behavior sample set and determine the missing state corresponding to each heterogeneous feature of each interaction behavior sample in the interaction behavior sample set. For each interaction behavior sample, the first interaction behavior feature sub-information corresponding to the interaction behavior sample is generated based on the feature values ​​and missing states of each heterogeneous feature in the interaction behavior sample. For each interaction behavior sample, based on the feature type and missing state of each heterogeneous feature in the interaction behavior sample, the heterogeneous features are embedded to generate the second interaction behavior feature sub-information corresponding to the interaction behavior sample. Based on the first interaction behavior feature sub-information corresponding to each interaction behavior sample, the first interaction behavior feature information corresponding to the interaction behavior sample set is determined, wherein the first interaction behavior feature sub-information includes the original feature value corresponding to the non-missing heterogeneous feature and the missing label corresponding to the missing heterogeneous feature. Based on the second interaction behavior feature sub-information corresponding to each interaction behavior sample, the second interaction behavior feature information corresponding to the interaction behavior sample set is determined. The second interaction behavior feature sub-information includes feature embedding information corresponding to heterogeneous features that are not missing and null value embedding information corresponding to missing heterogeneous features.

[0005] In one possible implementation, feature association pattern information between heterogeneous features is determined based on the second interaction behavior feature information, including: Multi-head attention processing is performed on the feature information of the second interaction behavior to obtain attention weight information; Based on attention weight information, determine the correlation strength information between any two heterogeneous features among multiple heterogeneous features; Based on the correlation strength information between any two heterogeneous features, feature association pattern information is generated.

[0006] In one possible implementation, missing features in the first interaction behavior feature information are completed based on feature association pattern information to obtain interaction behavior completion feature information, including: For each missing feature in the first interaction behavior feature information, the target association pattern information corresponding to the missing feature is obtained from the feature association pattern information; Based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature. Based on the completion guidance information corresponding to the missing features, the missing features are completed to obtain the completion results corresponding to the missing features. Based on the completion results corresponding to each missing feature in the first interaction behavior feature information and the non-missing features in the first interaction behavior feature information, interaction behavior completion feature information is generated.

[0007] In one possible implementation, based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature, including: When the missing feature is of continuous type, the first mapping network is used to map the target association pattern information corresponding to the missing feature to obtain the first completion guidance information for determining the distribution parameter corresponding to the missing feature. When the missing feature is of discrete type, a second mapping network is used to map the target association pattern information corresponding to the missing feature to obtain the second completion guidance information used to guide the information propagation in the heterogeneous graph. When the missing feature is of time-series type, a third mapping network is used to map the target association pattern information corresponding to the missing feature to obtain third completion guidance information for guiding time-dependent modeling.

[0008] In one possible implementation, based on the completion guidance information corresponding to the missing features, the missing features are completed to obtain the completed results corresponding to the missing features, including: When the missing feature is of continuous type, the distribution parameters corresponding to the missing feature are generated based on the corresponding completion guidance information, and the completion value corresponding to the missing feature is obtained by sampling from the probability distribution corresponding to the distribution parameters. The completion value is then determined as the completion result corresponding to the missing feature. When the missing feature is discrete, the neighbor node information in the heterogeneous graph is propagated and aggregated based on the corresponding completion guidance information to obtain the category probability information corresponding to the missing feature. The completion category corresponding to the missing feature is determined based on the category probability information, and the completion category is determined as the completion result corresponding to the missing feature. When the missing feature is of time-series type, time-dependency modeling is performed on the time-series data corresponding to the missing feature based on the corresponding completion guidance information to obtain the completion sequence corresponding to the missing feature, and the completion sequence is determined as the completion result corresponding to the missing feature.

[0009] In one possible implementation, based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set, the initial interaction behavior prediction model is trained to obtain the target interaction behavior prediction model, including: Determine the credibility information of the completion results corresponding to each missing feature in the feature information of interactive behavior completion; Based on the credibility information of the completion results corresponding to each missing feature, feature filtering processing is performed on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information. Input the target feature information of the interaction behavior into the initial interaction behavior prediction model to obtain the interaction behavior prediction results corresponding to each interaction behavior sample in the interaction behavior sample set. Based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

[0010] In one possible implementation, the reliability information of the completion result corresponding to each missing feature in the interaction behavior completion feature information is determined, including: When the missing feature is of continuous type, the confidence information corresponding to the missing feature is determined based on the variance information in the distribution parameters corresponding to the missing feature. When the missing feature is discrete, the confidence information corresponding to the missing feature is determined based on the maximum class probability in the class probability information corresponding to the missing feature. When the missing feature is of time-series type, the confidence information corresponding to the missing feature is determined based on at least one of the smoothness information and periodic consistency information of the completed sequence corresponding to the missing feature.

[0011] In one possible implementation, based on the credibility information of the completion results corresponding to each missing feature, feature filtering processing is performed on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information, including: The interaction behavior completion feature information is embedded to obtain the completion feature embedding information; Using the second interaction behavior feature information as query information and the completion feature embedding information as key and value information, cross-attention processing is performed on the second interaction behavior feature information and the completion feature embedding information to obtain cross-attention feature information. The second interaction behavior feature information and the cross attention feature information are subjected to feature cross processing to obtain interaction enhancement feature information; Based on the credibility information of the completion results corresponding to each missing feature, the gating weights corresponding to each feature in the interaction enhancement feature information are determined, and the interaction enhancement feature information is filtered based on the gating weights to obtain the interaction behavior target feature information.

[0012] In one possible implementation, based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model, including: Based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the interaction behavior prediction loss is determined. Based on the feature information corresponding to complete feature samples, missing feature samples and complete feature samples in the interaction behavior sample set, the contrastive learning loss is determined. Obtain attention weight information generated during the process of determining feature association pattern information, and determine attention regularization loss based on attention weight information; Based on the interaction behavior prediction loss, contrastive learning loss, and attention regularization loss, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

[0013] Secondly, embodiments of this application provide an interactive behavior prediction method, including: Acquire interactive behavior data to be predicted. The interactive behavior data to be predicted includes multiple heterogeneous features, including missing features. The interactive behavior data to be predicted is used to characterize the interaction scenario data between the target user and the target object. Input the interaction behavior data to be predicted into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be predicted. The interaction behavior prediction result is used to characterize the predicted probability of the target user performing the target interaction behavior on the target object. The target interaction behavior prediction model is generated by the interaction behavior prediction model training method provided in the first aspect or any possible implementation of the first aspect.

[0014] Thirdly, embodiments of this application provide an interactive behavior prediction model training device, including: The first acquisition module is used to acquire first interaction behavior feature information and second interaction behavior feature information corresponding to the interaction behavior sample set. The interaction behavior sample set includes multiple heterogeneous features. The first interaction behavior feature information corresponds to the original value space, and the second interaction behavior feature information corresponds to the embedding space. The determination module is used to determine the feature association pattern information between heterogeneous features based on the second interaction behavior feature information; The processing module is used to complete the missing features in the first interactive behavior feature information based on feature association pattern information to obtain interactive behavior completion feature information. The training module is used to train the initial interaction behavior prediction model based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set, so as to obtain the target interaction behavior prediction model.

[0015] Fourthly, embodiments of this application provide an interactive behavior prediction device, including: The second acquisition module is used to acquire the interaction behavior data to be estimated. The interaction behavior data to be estimated includes multiple heterogeneous features, including missing features. The interaction behavior data to be estimated is used to characterize the interaction scenario data between the target user and the target object. The inference module is used to input the interaction behavior data to be predicted into the target interaction behavior prediction model, and obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be predicted. The interaction behavior prediction result is used to characterize the predicted probability of the target user performing the target interaction behavior on the target object. The target interaction behavior prediction model is generated by the interaction behavior prediction model training method provided in the first aspect or any possible implementation of the first aspect.

[0016] Fifthly, embodiments of this application provide an electronic device, including: a processor and a memory; wherein the memory stores a computer program, and the processor executes the computer program to implement the method steps provided in the first or second aspect of embodiments of this application.

[0017] Sixthly, embodiments of this application provide a computer storage medium storing a plurality of instructions adapted for a processor to load and execute the method steps provided in the first or second aspect of embodiments of this application.

[0018] In a seventh aspect, embodiments of this application provide a computer program product containing instructions that, when run on a computer or processor, cause the computer or processor to execute the method provided in the first or second aspect of embodiments of this application.

[0019] In this embodiment, firstly, firstly, firstly, secondly, interactive behavior feature information and secondly, interactive behavior feature information corresponding to the interactive behavior sample set are obtained. The interactive behavior sample set includes multiple heterogeneous features. The firstly, interactive behavior feature information corresponds to the original value space, and the secondly, interactive behavior feature information corresponds to the embedding space. Subsequently, based on the secondly, interactive behavior feature information, feature association pattern information between heterogeneous features is determined. Then, based on the feature association pattern information, missing features in the firstly, interactive behavior feature information are completed to obtain interactive behavior completion feature information. Finally, based on the interactive behavior completion feature information and the interactive behavior labels corresponding to the interactive behavior sample set, the initial interactive behavior prediction model is trained to obtain the target interactive behavior prediction model. Therefore, by constructing the first interactive behavior feature information in the original value space and the second interactive behavior feature information in the embedding space respectively, it is possible to preserve the original value states of each heterogeneous feature while utilizing the semantic information in the embedding space to mine the correlation between different heterogeneous features. Furthermore, based on the feature association pattern information, the missing features in the original value space are completed, so that the completion results can refer to the information of other related features, thereby improving the rationality and accuracy of missing feature completion. On this basis, the interactive behavior prediction model is trained using the completed interactive behavior feature information and the corresponding interactive behavior labels, which can reduce the impact of feature missingness on the model training process and prediction results, thereby improving the prediction accuracy, stability and generalization ability of the target interactive behavior prediction model in feature missing scenarios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 An exemplary system architecture diagram of an interactive behavior prediction model training method provided in this application embodiment; Figure 2 A flowchart illustrating a training method for an interactive behavior prediction model provided in an embodiment of this application; Figure 3 A flowchart illustrating a method for determining interactive behavior feature information provided in an embodiment of this application; Figure 4 A flowchart illustrating a method for determining feature association pattern information provided in an embodiment of this application; Figure 5 A flowchart illustrating a method for determining interactive behavior completion feature information provided in an embodiment of this application; Figure 6A flowchart illustrating a method for generating a target interaction behavior prediction model, provided in an embodiment of this application; Figure 7 A schematic diagram of the overall process of training an interactive behavior prediction model provided in this application embodiment; Figure 8 A schematic diagram of the structure of an interactive behavior prediction model training device provided in an embodiment of this application; Figure 9 A schematic diagram of the structure of an interactive behavior prediction device provided in an embodiment of this application; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0022] To make the features and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this application.

[0023] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims. Furthermore, in the description of the embodiments of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B; the word "and / or" in the text is merely a description of the association relationship between related objects, indicating that three relationships can exist, for example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Additionally, in the description of the embodiments of this application, "multiple" refers to two or more.

[0024] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0025] In applications such as product recommendation, content recommendation, and advertising, it is often necessary to predict the probability of a user performing interactive behaviors such as clicking on a recommended target based on multiple user characteristics. However, due to factors such as sparse user behavior, incomplete data collection, and untimely data synchronization across systems, the feature information input to the model may be missing to varying degrees. When the proportion of missing features is high, the interactive behavior prediction models in related technologies struggle to adapt to the differences in data structure and variation patterns of different types of features, resulting in low accuracy of missing feature processing results. This further affects the accuracy, stability, and generalization ability of the interactive behavior prediction results.

[0026] For example, in content recommendation scenarios, newly registered users may have limited or missing features such as browsing history, content preferences, and interaction frequency, making it difficult for interaction behavior prediction models to accurately determine the probability of a user clicking on recommended content. Similarly, in consumer finance scenarios, user applications, credit granting, borrowing, and repayments are infrequent and cyclical, and some user characteristics may be missing due to data collection conditions, user authorization status, or cross-system data synchronization issues. This makes it difficult for models to obtain complete user attributes, historical behavior, and credit-related features, thus affecting the predicted probability of users clicking on loan products, credit card products, and other target objects.

[0027] Therefore, this application provides a training method for an interactive behavior prediction model to solve the technical problems of weak generalization ability and low accuracy of prediction results of the aforementioned interactive behavior prediction model.

[0028] Please see Figure 1 , Figure 1 An exemplary system architecture diagram of an interactive behavior prediction model training method provided in this application embodiment.

[0029] like Figure 1 As shown, the system architecture may include a terminal 101, a network 102, and a server 103. The network 102 serves as the medium for providing a communication link between the terminal 101 and the server 103. The network 102 may include various types of wired or wireless communication links, such as wired communication links including fiber optic cables, twisted-pair cables, or coaxial cables, and wireless communication links including Bluetooth communication links, Wireless-Fidelity (Wi-Fi) communication links, or microwave communication links, etc.

[0030] Terminal 101 can interact with server 103 via network 102 to receive messages from or send messages to server 103. Alternatively, terminal 101 can interact with server 103 via network 102 to receive messages or data sent to server 103 by other users. Terminal 101 can be hardware or software. When terminal 101 is hardware, it can be various electronic devices, including but not limited to tablet computers, laptops, and desktop computers. When terminal 101 is software, it can be installed in the aforementioned electronic devices and can be implemented as multiple software programs or software modules (e.g., to provide distributed services) or as a single software program or software module; no specific limitation is made here.

[0031] In this application, during the training preparation phase, server 103 can acquire first and second interactive behavior feature information corresponding to the interactive behavior sample set. The interactive behavior sample set includes multiple heterogeneous features. The first interactive behavior feature information corresponds to the original value space, and the second interactive behavior feature information corresponds to the embedding space. Subsequently, server 103 determines the feature association pattern information between heterogeneous features based on the second interactive behavior feature information. Based on the feature association pattern information, server 103 performs feature completion processing on the missing features in the first interactive behavior feature information to obtain interactive behavior completion feature information. Then, based on the interactive behavior completion feature information and the interactive behavior labels corresponding to the interactive behavior sample set, server 103 trains the initial interactive behavior prediction model to obtain the target interactive behavior prediction model.

[0032] Furthermore, server 103 can obtain the interaction behavior data to be predicted from terminal 101 via network 102. This data includes multiple heterogeneous features, including missing features. The interaction behavior data is used to characterize the interaction scenario data between the target user and the target object. Subsequently, server 103 can input the interaction behavior data into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the model. This result characterizes the predicted probability that the target user will perform a target interaction behavior towards the target object. Optionally, server 103 can display the interaction behavior prediction result to relevant users for them to perform related downstream processing.

[0033] For example, taking a product recommendation scenario, terminal 101 can be a shopping terminal used by a user, the target object can be the target product to be recommended by server 103, and the target interaction behavior can be the click behavior performed by the user on the target product. Server 103 can obtain an interaction behavior sample set including multiple historical user samples. Each historical user sample can include user attribute features, historical browsing features, historical purchase features, target product attribute features, and interaction tags between the user and the target product. Among them, some historical user samples may have missing features such as historical browsing count, product preference category, or purchase frequency due to reasons such as short user registration time, few historical behaviors, or incomplete data synchronization. During the model training phase, server 103 can generate first interaction behavior feature information in the original value space and second interaction behavior feature information in the embedding space of the interaction behavior sample set, and determine the feature association pattern information between different features based on the second interaction behavior feature information, and complete the missing features such as product preference category or purchase frequency based on the association relationship. Server 103 further trains the initial interaction behavior prediction model based on the completed interaction behavior feature information and the click tags corresponding to each historical user sample to obtain the target interaction behavior prediction model. During the model application phase, terminal 101 can respond to a user's entry into the product recommendation page by sending the current user's user characteristics, historical behavior characteristics, target product characteristics, and scenario characteristics—all pending prediction of interactive behavior data—to server 103. When missing features exist in the pending prediction data, server 103 can input the pending prediction data into the target interactive behavior prediction model. The target interactive behavior prediction model processes the missing features and outputs the predicted probability of the current user clicking on the target product. Furthermore, server 103 can determine the recommendation order of target products based on the predicted probabilities and send the corresponding product recommendation results to terminal 101 for display. For example, if the predicted probability of a target product is higher than the predicted probabilities of other candidate products, the target product can be placed in a more prominent position on the product recommendation page.

[0034] Optionally, server 103 can be a server that provides various services. It should be noted that server 103 can be hardware or software. When server 103 is hardware, it can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When server 103 is software, it can be implemented as multiple software programs or software modules (e.g., used to provide distributed services), or as a single software program or software module; no specific limitations are made here.

[0035] It should be understood that Figure 1 The number of terminals, networks, and servers shown is only illustrative; the number can be any number of terminals, networks, and servers depending on the implementation requirements.

[0036] Please see Figure 2 , Figure 2 This is a flowchart illustrating a training method for an interactive behavior prediction model provided in an embodiment of this application. Figure 2 As shown, the training methods for interactive behavior prediction models can include at least: S202: Obtain the first interaction behavior feature information and the second interaction behavior feature information corresponding to the interaction behavior sample set. The interaction behavior sample set includes multiple heterogeneous features. The first interaction behavior feature information corresponds to the original value space, and the second interaction behavior feature information corresponds to the embedding space.

[0037] The interaction behavior sample set can include multiple interaction behavior samples, each of which can be used to characterize historical interaction scenarios between the target user and the target object. Optionally, the target object can be a product, content, advertisement, service, or other interactive object to be recommended, and the interaction behavior can include actions such as clicking, browsing, favoriting, applying, or submitting. Each interaction behavior sample can include user characteristics, target object characteristics, historical behavior characteristics, and scenario characteristics, as well as interaction behavior tags used to characterize whether the user performed the target interaction behavior.

[0038] It is understandable that heterogeneous features can refer to features with different data types, data structures, or patterns of change, such as continuous features, discrete features, and time-series features. In real-world scenarios, due to factors such as limited user historical behavior, incomplete data collection, or data synchronization delays, some features in each interaction behavior sample may be missing.

[0039] It should be noted that the first interaction behavior feature information is used to record the original feature values ​​and missing states of each heterogeneous feature in the original value space. For example, for non-missing features, the corresponding original feature values ​​are recorded, and for missing features, the corresponding missing labels are recorded. The second interaction behavior feature information is used to represent the semantic features of each heterogeneous feature in the embedding space. For example, non-missing features can be embedded, and missing features can be represented using null value embedding information.

[0040] S204: Based on the second interaction behavior feature information, determine the feature association pattern information between heterogeneous features.

[0041] Among them, feature association pattern information can be used to characterize the association relationship and degree of association between different heterogeneous features.

[0042] Optionally, feature association analysis can be performed on the second interaction behavior feature information to determine the association strength information between the heterogeneous features, and feature association pattern information can be generated based on the association strength information between the heterogeneous features. For example, the semantic association relationship between user attribute features, historical behavior features, and target object features can be determined.

[0043] S206: Based on the feature association pattern information, the missing features in the first interaction behavior feature information are completed to obtain the interaction behavior completion feature information.

[0044] The interactive behavior completion feature information can include the completion results for each missing feature and the original feature values ​​for each non-missing feature. For different types of missing features, the completion results can be expressed as completion values, completion categories, or completion sequences.

[0045] Optionally, information can be determined to guide the completion of the missing feature based on the relationship between the missing feature and other features, and a completion result can be generated using a processing method that is compatible with the feature type of the missing feature.

[0046] S208: Based on the feature information of the interaction behavior completion and the interaction behavior labels corresponding to the interaction behavior sample set, the initial interaction behavior prediction model is trained to obtain the target interaction behavior prediction model.

[0047] The interaction behavior labels corresponding to the interaction behavior sample set can be label information that corresponds to each interaction behavior sample in the set and is used to characterize whether the target user performs the target interaction behavior towards the target object. For example, in a product recommendation scenario, the interaction behavior label can be a click label to characterize whether the user clicks on the product to be recommended; in a content recommendation scenario, the interaction behavior label can be used to characterize whether the user clicks on, browses, or favorites the content to be recommended. The interaction behavior labels can be binary labels, for example, using a first preset value to indicate that the target interaction behavior has occurred and a second preset value to indicate that the target interaction behavior has not occurred; other label forms can also be used according to the actual task.

[0048] It should be noted that the initial interaction behavior prediction model refers to the interaction behavior prediction model that has not yet completed this round of model training, or whose model parameters have not yet been adjusted to meet the preset training termination conditions. The target interaction behavior prediction model refers to the model obtained after training the initial interaction behavior prediction model based on the interaction behavior completion feature information and corresponding interaction behavior labels, and meeting the preset training termination conditions. The target interaction behavior prediction model can be used to output the predicted probability of a target user performing a target interaction behavior on a target object based on the interaction behavior data to be predicted. The preset training termination conditions may include at least one of the following: model training loss convergence, the number of model training iterations reaching a preset number, and the model's prediction performance on the validation sample set meeting preset requirements.

[0049] In this embodiment, firstly, firstly, firstly, secondly, interactive behavior feature information and secondly, interactive behavior feature information corresponding to the interactive behavior sample set are obtained. The interactive behavior sample set includes multiple heterogeneous features. The firstly, interactive behavior feature information corresponds to the original value space, and the secondly, interactive behavior feature information corresponds to the embedding space. Subsequently, based on the secondly, interactive behavior feature information, feature association pattern information between heterogeneous features is determined. Then, based on the feature association pattern information, missing features in the firstly, interactive behavior feature information are completed to obtain interactive behavior completion feature information. Finally, based on the interactive behavior completion feature information and the interactive behavior labels corresponding to the interactive behavior sample set, the initial interactive behavior prediction model is trained to obtain the target interactive behavior prediction model. Therefore, by constructing the first interactive behavior feature information in the original value space and the second interactive behavior feature information in the embedding space respectively, it is possible to preserve the original value states of each heterogeneous feature while utilizing the semantic information in the embedding space to mine the correlation between different heterogeneous features. Furthermore, based on the feature association pattern information, the missing features in the original value space are completed, so that the completion results can refer to the information of other related features, thereby improving the rationality and accuracy of missing feature completion. On this basis, the interactive behavior prediction model is trained using the completed interactive behavior feature information and the corresponding interactive behavior labels, which can reduce the impact of feature missingness on the model training process and prediction results, thereby improving the prediction accuracy, stability and generalization ability of the target interactive behavior prediction model in feature missing scenarios.

[0050] In one embodiment, see Figure 3 In step S202 above, obtaining the first interaction behavior feature information and the second interaction behavior feature information corresponding to the interaction behavior sample set may include the following steps: S302: Obtain the interaction behavior sample set and determine the missing state corresponding to each heterogeneous feature of each interaction behavior sample in the interaction behavior sample set.

[0051] Each interaction behavior sample in the interaction behavior sample set can be used to characterize the feature state and interaction result of a target user and a target object in a corresponding interaction scenario. For example, in a product recommendation scenario, an interaction behavior sample can correspond to an exposure event between a user and a product to be recommended; in a content recommendation scenario, an interaction behavior sample can correspond to a recommendation event between a user and content to be recommended; and in a consumer finance product recommendation scenario, an interaction behavior sample can correspond to a display event between a user and a financial product.

[0052] Optionally, the heterogeneous features of each interaction behavior sample can refer to multiple features in each interaction behavior sample that have different data types, data structures, or change patterns. For example, heterogeneous features may include at least one of the following: basic user information features, such as occupation category; credit history features, such as credit score, number of overdue records, and number of historical loans; behavioral features, such as login frequency in the past 30 days, number of visits to the target object page in the past 7 days, and the proportion of historical clicked links; time-series features, such as recent transaction frequency sequences, behavioral regularity sequences, and resource usage cycle pattern sequences; third-party features obtained with user authorization, such as third-party credit information, device profile information, or user preference information; target object features, such as product category, content category, advertising type, service type, and target object attributes; and scenario features, such as recommendation time, recommendation location, terminal type, and access channel.

[0053] It should be noted that heterogeneous features can include continuous features, discrete features, and time-series features. Continuous features are those whose feature values ​​are typically continuous numerical values, such as credit scores and visit counts. Discrete features are those whose feature values ​​are selected from a predefined set of categories, such as occupation categories, product categories, and channel categories. Time-series features are sequences of one or more feature values ​​arranged in chronological order, such as a sequence of visit counts over recent days, a transaction frequency sequence, or a behavioral cycle sequence.

[0054] It is understandable that feature values ​​can refer to the specific values ​​of corresponding heterogeneous features in the corresponding interaction behavior samples.

[0055] In real-world applications, due to factors such as limited user history, the target object appearing for the first time, unauthorized data collection, data collection failure, cross-system synchronization delays, or the incomplete execution of certain service processes, some heterogeneous features in the interaction behavior samples may lack valid feature values. In such cases, the corresponding heterogeneous features can be identified as missing features.

[0056] Optionally, the missing state can be used to characterize whether a valid feature value exists for the corresponding heterogeneous feature. For the i-th dimension heterogeneous feature, a missing indicator bit m corresponding to that heterogeneous feature can be generated. i If the eigenvalue of the i-th heterogeneous feature exists, then the corresponding missing indicator bit m is set. i Set to 1; if the eigenvalue of the i-th heterogeneous feature is missing, then set the corresponding missing value indicator m. i Set to 0, that is: m i =1, the eigenvalue of the i-th heterogeneous feature exists; m i =0, the eigenvalue of the i-th heterogeneous feature is missing.

[0057] Furthermore, based on the missing indicator bits corresponding to each heterogeneous feature in each interaction behavior sample, a binary mask vector corresponding to that interaction behavior sample can be generated: Where F represents the total number of features included in the interactive behavior sample, and the i-th element in M ​​is used to represent the missing state of the i-th heterogeneous feature.

[0058] In some implementations, the missing status of each heterogeneous feature can be determined by checking whether the feature field is empty, whether the feature value is a preset invalid value, whether the feature exceeds the valid time range, and whether the corresponding data interface returns valid data. For example, in a product recommendation scenario, if the user's registration time is short and the historical purchase frequency field has not yet been generated, it can be determined that the historical purchase frequency feature does not have a valid feature value; if the user's occupation category field is empty, or the value of this field is a preset invalid character, it can be determined that the occupation category feature does not have a valid feature value; if the data update time corresponding to the number of product visits in the past 30 days is earlier than the current time preset duration, causing the feature to exceed the valid time range, the number of product visits in the past 30 days feature can be determined as a missing feature. For example, in consumer finance product recommendation scenarios, if a user has not yet borrowed or repaid any loans, features such as the number of historical loans and repayment patterns may not yet be generated, and these features can be identified as missing features. If the user has not authorized access to third-party credit information, or if the third-party data interface has not returned valid data, the corresponding third-party credit features can be identified as missing features. If the data between the credit granting system and the recommendation system has not yet been synchronized, credit-related features for which valid values ​​cannot be obtained at present can be identified as missing features. Based on the above methods, the missing status of each heterogeneous feature in each interaction behavior sample can be determined separately, and corresponding missing indicator bits can be generated.

[0059] S304: For each interaction behavior sample, based on the feature values ​​and missing states of each heterogeneous feature in the interaction behavior sample, generate the first interaction behavior feature sub-information corresponding to the interaction behavior sample.

[0060] The first interactive behavior feature sub-information refers to the feature information of a single interactive behavior sample in the original value space, used to preserve the original value state and missing state of each heterogeneous feature in the interactive behavior sample. The original value space refers to the space in which feature information is recorded according to the original data form of each heterogeneous feature. In the original value space, heterogeneous features that are not missing can retain their original feature values, and missing heterogeneous features can be marked with missing markers.

[0061] Specifically, for a given interaction behavior sample, the heterogeneous features in the sample can be arranged according to a preset feature arrangement order. For heterogeneous features indicating the presence of a missing value, the corresponding original feature value is written to the corresponding feature position; for heterogeneous features indicating a missing value, a missing marker is written to the corresponding feature position. For example, the missing marker can be a null value (NULL), a preset character, a preset numerical value, or other markers that can distinguish it from a valid feature value.

[0062] In the embodiments of this application, the missing marker can be used to indicate that there is no valid original feature value at the corresponding feature position, and does not mean that the missing marker is the final completion value of the missing feature.

[0063] S306: For each interaction behavior sample, based on the feature type and missing state of each heterogeneous feature in the interaction behavior sample, embedding processing is performed on each heterogeneous feature to generate the second interaction behavior feature sub-information corresponding to the interaction behavior sample.

[0064] The second interactive behavior feature sub-information refers to the feature information of a single interactive behavior sample in the embedding space, used to represent the semantic information of different types of heterogeneous features in a vector form with a unified dimension. The embedding space refers to the feature space formed after heterogeneous features of different data types, different value ranges, or different data structures are converted into vector information.

[0065] Optionally, the feature type can be used to characterize the data form corresponding to the heterogeneous features, such as continuous, discrete, and time-series data. The appropriate embedding method can be selected based on the feature type of the heterogeneous features.

[0066] For example, for continuous features that are not missing, the original feature values ​​of the continuous features can be mapped to D-dimensional vectors using a linear projection layer, and the D-dimensional vectors are determined as the feature embedding information corresponding to the continuous features. Here, D represents the embedding dimension. For discrete features that are not missing, the class index of the discrete features can be determined, and the class index can be mapped to a D-dimensional vector using an embedding lookup table, and the D-dimensional vectors are determined as the feature embedding information corresponding to the discrete features. For heterogeneous features that are missing, learnable null embedding vectors can be used to represent the missing state of the heterogeneous features in the embedding space, and the learnable null embedding vectors are determined as the null embedding information corresponding to the missing features.

[0067] Therefore, for an interaction behavior sample, based on the feature type and missing state of each heterogeneous feature in the interaction behavior sample, the feature embedding information corresponding to the non-missing heterogeneous features and the null value embedding information corresponding to the missing heterogeneous features can be obtained respectively, and the second interaction behavior feature sub-information corresponding to the interaction behavior sample can be generated based on the feature embedding information and the null value embedding information.

[0068] S308: Based on the first interaction behavior feature sub-information corresponding to each interaction behavior sample, determine the first interaction behavior feature information corresponding to the interaction behavior sample set, wherein the first interaction behavior feature sub-information includes the original feature value corresponding to the non-missing heterogeneous feature and the missing label corresponding to the missing heterogeneous feature.

[0069] Optionally, the first interaction behavior feature sub-information corresponding to each interaction behavior sample can be combined according to the order of each interaction behavior sample in the interaction behavior sample set to obtain the first interaction behavior feature information. For example, the first interaction behavior feature information can be represented as an original feature matrix: Where N represents the number of interaction behavior samples in the interaction behavior sample set, and F represents the total number of features corresponding to each interaction behavior sample. Original feature matrix The nth row can correspond to the first interaction behavior feature sub-information of the nth interaction behavior sample, and the ith column can correspond to the ith-dimensional heterogeneous feature.

[0070] For the original feature matrix elements in If the i-th heterogeneous feature of the nth interaction behavior sample is not missing, then This can be the original feature value of the heterogeneous feature; if the i-th dimension of the heterogeneous feature of the nth interaction behavior sample is missing, then NULL can be used to mark missing values.

[0071] It should be noted that, since the original feature values ​​of discrete features can be character categories or category indices, the original feature matrix... Logically, it is used to represent the correspondence between N samples and F feature dimensions, but it does not require all matrix elements to use the same data storage type. In actual implementation, different feature types can be stored separately, or they can be stored uniformly after format conversion.

[0072] S310: Based on the second interaction behavior feature sub-information corresponding to each interaction behavior sample, determine the second interaction behavior feature information corresponding to the interaction behavior sample set, wherein the second interaction behavior feature sub-information includes feature embedding information corresponding to the heterogeneous features that are not missing and null value embedding information corresponding to the missing heterogeneous features.

[0073] Optionally, the second interactive behavior feature sub-information corresponding to each interactive behavior sample can be combined according to the arrangement order of each interactive behavior sample and each heterogeneous feature to obtain the second interactive behavior feature information.

[0074] For example, the second interaction behavior feature information can be represented as an embedded feature tensor: Where N represents the number of interaction behavior samples, F represents the total number of features, and D represents the embedding dimension.

[0075] In this embodiment, by generating first interactive behavior feature sub-information in the original value space and second interactive behavior feature sub-information in the embedding space for each interactive behavior sample, and further combining them to obtain the first and second interactive behavior feature information corresponding to the interactive behavior sample set, it is possible to convert heterogeneous features of different types into embedding information of a unified dimension while preserving the original values ​​of the features that are not missing and the positional relationships of the missing features. The first interactive behavior feature information can provide the original feature basis for subsequent missing feature completion and backfilling of completion results, while the second interactive behavior feature information can provide a unified data basis for subsequent mining of semantic relationships between different heterogeneous features. Through the above dual-space feature processing method, the impact of different feature types, value ranges, and missing states on the subsequent model processing can be reduced, and the accuracy and stability of missing feature completion and interactive behavior prediction can be improved.

[0076] In one embodiment, see Figure 4 In S204, determining the feature association pattern information between heterogeneous features based on the second interaction behavior feature information may include the following steps: S402: Perform multi-head attention processing on the second interaction behavior feature information to obtain attention weight information.

[0077] The attention weight information can be used to characterize the degree of attention between different heterogeneous features in each interaction behavior sample. Each element in the attention weight information can be used to characterize the degree of attention one heterogeneous feature pays to another under the h-th attention head. The larger the value of the corresponding element, the higher the degree of attention between the two heterogeneous features; the smaller the value of the corresponding element, the lower the degree of attention between the two heterogeneous features.

[0078] Understandably, different attention heads can focus on different types of feature relationships. For example, one attention head may focus primarily on the relationship between user attribute features and historical behavior features, another attention head may focus primarily on the relationship between historical behavior features and target object features, and yet another attention head may focus primarily on the periodic or trend-based relationships between temporal features.

[0079] For example, the second interaction behavior feature information can be... The input is a multi-head self-attention network, which can include H parallel attention heads. For the h-th attention head, the query matrix Q can be calculated separately. h Key matrix K h Sum matrix V h : , , .in, , and Let be the learnable projection matrix corresponding to the h-th attention head, and: , , ,in, This represents the vector dimension corresponding to a single attention head. In one optional implementation, It can be set to D / H, that is, the original embedding dimension D is divided into H attention heads.

[0080] It should be noted that the query matrix It can be used to characterize the association information that needs to be obtained from various heterogeneous features under the current attention head; key matrix It can be used to characterize the matching information that various heterogeneous features can provide; value matrix It can be used to characterize the feature content provided by various heterogeneous features after establishing association.

[0081] Furthermore, it can be based on the query matrix Bond matrix Determine the attention weight corresponding to the h-th attention head. The attention weight corresponding to the h-th attention head can be expressed as: ;in, Used to characterize the degree of matching between different heterogeneous features. This is used to scale the matching results to avoid excessively large dot product results due to large vector dimensions. This is used to convert the scaled matching results into normalized attention weights. Optionally, It can be set to D / H, that is, the ratio of the embedding dimension of the second interaction behavior feature information to the number of attention heads.

[0082] It is understandable that, for any interaction behavior sample in the interaction behavior sample set, each element in the attention weight corresponding to the h-th attention head can be used to characterize the degree of attention one heterogeneous feature pays to another heterogeneous feature under the h-th attention head. The attention weights corresponding to each attention head can form attention weight information.

[0083] Furthermore, the attention outputs corresponding to each attention head can be concatenated and a comprehensive attention representation can be obtained through linear projection, which can be specifically represented as: ;in, This indicates splicing / joining. This represents the output projection matrix, and .

[0084] Therefore, the relationship between different heterogeneous features can be modeled in the embedding space by using multiple parallel attention heads, and the attention weight information corresponding to each attention head can be obtained.

[0085] S404: Based on attention weight information, determine the correlation strength information between any two heterogeneous features among multiple heterogeneous features.

[0086] The association strength information between any two heterogeneous features refers to the degree of semantic association between one heterogeneous feature and another in the corresponding interaction behavior sample. The association strength information can be determined based on the attention weight information. The larger the value of the association strength information, the stronger the semantic association between the two corresponding heterogeneous features; the smaller the value of the association strength information, the weaker the semantic association between the two corresponding heterogeneous features.

[0087] Specifically, for the i-th and j-th heterogeneous features in the n-th interaction behavior sample, the attention weights corresponding to features i and j can be obtained from the attention weight information, and the attention weights can be determined as or converted into the association strength information between features i and j.

[0088] In one implementation, the attention weights corresponding to features i and j in a specific attention head can be directly used as the association strength information between features i and j. In another implementation, aggregation processing can be performed based on the attention weights corresponding to features i and j in multiple attention heads, such as summation, averaging, maximum value processing, or weighted fusion processing, to obtain the comprehensive association strength information between features i and j.

[0089] For example, the association strength information between the i-th heterogeneous feature and the j-th heterogeneous feature in the n-th interaction behavior sample can be represented as: .

[0090] It should be noted that, and They can be the same or different. When using directional attention weights, It can represent the degree of attention that feature i pays to feature j. This can represent the degree of attention that feature j pays to feature i. When it is necessary to construct a non-directional association, the two can also be averaged or subjected to other symmetric processing.

[0091] S406: Generate feature association pattern information based on the association strength information between any two heterogeneous features.

[0092] Feature association pattern information can be used to comprehensively characterize the relationships and strengths of associations among multiple heterogeneous features in each interaction behavior sample. Feature association pattern information can include the feature association matrix corresponding to each interaction behavior sample, or it can include a feature association tensor formed by the feature association matrices corresponding to multiple interaction behavior samples.

[0093] Specifically, the correlation strength information between any two heterogeneous features in each interaction behavior sample can be extracted from the attention weight information, and then combined according to the interaction behavior sample dimension and the heterogeneous feature dimension to obtain the feature association pattern information P corresponding to the interaction behavior sample set. The feature association pattern information can be represented as: Where N represents the number of interaction behavior samples in the interaction behavior sample set, and F represents the number of heterogeneous features corresponding to each interaction behavior sample.

[0094] It should be noted that feature association pattern information is extracted from the attention mechanism and is used to characterize the semantic relationships between different heterogeneous features. For any missing feature, the association strength information between the missing feature and other heterogeneous features can be obtained from the feature association matrix of the corresponding interaction behavior sample based on the feature dimension corresponding to the missing feature, so as to obtain the target association pattern information corresponding to the missing feature, which is then used to generate the corresponding completion guidance information.

[0095] In this embodiment, by performing multi-head attention processing on the second interaction behavior feature information, attention relationships between different heterogeneous features can be mined from multiple feature subspaces, obtaining attention weight information that can characterize the degree of attention between each heterogeneous feature. Furthermore, based on the attention weight information, the association strength information between any two heterogeneous features is determined, and the association strength information between multiple heterogeneous features in each interaction behavior sample is combined to generate structured feature association pattern information. This allows for a more comprehensive capture of the potential semantic relationships between different types of features, and records the degree of association between each heterogeneous feature in the form of feature association pattern information. This provides a basis for subsequent extraction of relevant feature information for missing features, generation of completion guidance information, and improvement of the accuracy of missing feature completion.

[0096] In one embodiment, see Figure 5 In S206, the missing features in the first interaction behavior feature information are completed based on feature association pattern information to obtain interaction behavior completion feature information, which may include the following steps: S502: For each missing feature in the first interactive behavior feature information, obtain the target association pattern information corresponding to the missing feature from the feature association pattern information.

[0097] Among them, the target association pattern information can be information extracted for the missing feature to be completed, which is used to characterize the association relationship and the strength of the association between the missing feature and other heterogeneous features.

[0098] Specifically, the feature association pattern information can be denoted as P. For the i-th missing feature in the n-th interaction behavior sample, the association strength information corresponding to the i-th feature dimension can be extracted from the feature association pattern information P according to the feature dimension i corresponding to the missing feature, so as to obtain the target association pattern information corresponding to the i-th missing feature.

[0099] S504: Based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature.

[0100] The completion guidance information can be used to control or assist the corresponding feature completion model in performing the completion process. The completion guidance information can be a completion guidance vector, which can be denoted as... Its dimension can be G.

[0101] Optionally, the completion guidance information may include, but is not limited to: first completion guidance information generated for continuous missing features, second completion guidance information generated for discrete missing features, and third completion guidance information generated for time-series missing features.

[0102] In one embodiment, in S504, based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature. This may include the following steps: when the feature type of the missing feature is continuous, a first mapping network is used to map the target association pattern information corresponding to the missing feature to obtain first completion guidance information for determining the distribution parameters corresponding to the missing feature; when the feature type of the missing feature is discrete, a second mapping network is used to map the target association pattern information corresponding to the missing feature to obtain second completion guidance information for guiding information propagation in heterogeneous graphs; when the feature type of the missing feature is temporal, a third mapping network is used to map the target association pattern information corresponding to the missing feature to obtain third completion guidance information for guiding time-dependent modeling.

[0103] Specifically, the mapping network (first mapping network / second mapping network / third mapping network) can map the target association pattern information corresponding to the i-th missing feature into a completion guidance vector. This mapping process can convert the feature association patterns obtained in the embedding space into guidance information suitable for the completion of the original value space, thereby establishing a mapping relationship between the semantic association information of the embedding space and the completion of missing features in the original value space.

[0104] In one embodiment, the first mapping network can be a multilayer perceptron network for processing continuous feature association pattern information, or it can be a continuous feature mapping network. The first mapping network may include one or more fully connected layers and corresponding activation function layers, used to convert the target association pattern information into a completion guidance vector suitable for a continuous feature probabilistic completion model. For the i-th continuous missing feature, the mapping process of the first mapping network can be expressed as: .in, This represents the first completion guidance vector corresponding to the i-th continuous missing feature. Indicates the first mapping network, This represents the target association pattern information corresponding to the i-th missing feature.

[0105] Optionally, the first completion guidance information output by the first mapping network may include information needed to subsequently determine the distribution parameters of the continuous missing features. The distribution parameters may refer to parameters used to describe the probability distribution of candidate values ​​for the missing features, such as the mean parameter and the variance parameter.

[0106] In one embodiment, the second mapping network can be a multilayer perceptron network for processing discrete feature association pattern information, or it can be a discrete feature mapping network. The second mapping network is used to convert the target association pattern information corresponding to discrete missing features into a completion guidance vector suitable for graph neural networks or other relational reasoning models. For the i-th discrete missing feature, the mapping process of the second mapping network can be represented as follows: .in, This represents the second completion guidance vector corresponding to the i-th discrete missing feature. Indicates the second mapping network, This represents the target association pattern information corresponding to the i-th missing feature. The heterogeneous graph can be a graph structure including multiple node types and / or multiple edge types. Optionally, the heterogeneous graph can include user nodes, feature nodes, and target object nodes. User nodes are used to represent target users or historical users; feature nodes are used to represent user attribute features, historical behavior features, target object attribute features, etc.; target object nodes are used to represent products, content, advertisements, services, or other objects to be recommended. Edges in the heterogeneous graph can be used to represent the association relationships between different nodes. For example, the edge between a user node and a feature node can be used to represent that the user has the corresponding feature; the edge between a user node and a target object node can be used to represent that the user has engaged in interactive behaviors such as browsing, clicking, favoriting, or purchasing the target object; and the edge between a feature node and a target object node can be used to represent the association relationship between the corresponding feature and the target object.

[0107] In one embodiment, the third mapping network can be a multilayer perceptron network for processing temporal feature association pattern information, or it can be a temporal feature mapping network. The third mapping network is used to convert the target association pattern information corresponding to the temporal missing feature into a completion guidance vector suitable for the temporal model. For the i-th temporal missing feature, the mapping process of the third mapping network can be represented as: ;in, This represents the completion guidance vector corresponding to the i-th time-series missing feature. Indicates the third mapping network, This represents the target association pattern information corresponding to the i-th missing feature. Specifically, time dependency modeling refers to modeling the short-term changes, long-term trends, periodicity, or seasonality of time-series data based on the dependencies between different time positions. Time dependency modeling can be implemented using temporal convolutional networks, recurrent neural networks, attention networks, or other time-series models.

[0108] Optionally, the third completion guidance information can be used to determine the degree of attention that a time series model pays to different time positions, time spans, or time patterns when processing time series data. For example, the completion guidance information can be converted into time attention weights to guide the time series model to focus on historical time positions that are highly correlated with the missing time positions; it can also be used to adjust the convolution processing or feature aggregation process in a time series convolutional model. For example, for a missing access frequency sequence of the last 30 days, completion guidance information can be generated based on the user's historical click ratio, target object category, recent login frequency, and the correlation patterns between other time series features. Based on this completion guidance information, time-dependent modeling of the periodicity, trend, and local changes in the access frequency sequence can be performed to provide a basis for the subsequent generation of the completed sequence.

[0109] It should be noted that the first, second, and third mapping networks can have the same or different network structures, but their network parameters can be independent of each other. Since different mapping networks are trained for different types of missing features, each mapping network can learn a mapping relationship that matches the corresponding feature type.

[0110] In some implementations, a correspondence between feature types and mapping networks can be established in advance based on the feature types of each heterogeneous feature. After determining the feature type of the currently missing feature, a first mapping network, a second mapping network, or a third mapping network can be selected according to the correspondence to perform mapping processing on the target association pattern information corresponding to the missing feature.

[0111] Therefore, by selecting the appropriate mapping network according to the feature type of the missing feature, the feature association pattern in the embedding space can be converted into completion guidance information that matches the continuous, discrete or temporal feature completion process, so that the subsequent feature completion processing can adapt to the data structure and change pattern of different types of features.

[0112] S506: Based on the completion guidance information corresponding to the missing features, the missing features are completed to obtain the completion results corresponding to the missing features.

[0113] In one embodiment, in S506, the missing feature is completed based on the completion guidance information corresponding to the missing feature to obtain the completion result corresponding to the missing feature. This may include the following steps: if the feature type of the missing feature is continuous, a distribution parameter corresponding to the missing feature is generated based on the corresponding completion guidance information, and samples are taken from the probability distribution corresponding to the distribution parameter to obtain the completion value corresponding to the missing feature, and the completion value is determined as the completion result corresponding to the missing feature; if the feature type of the missing feature is discrete, the neighbor node information in the heterogeneous graph is propagated and aggregated based on the corresponding completion guidance information to obtain the category probability information corresponding to the missing feature, and the completion category corresponding to the missing feature is determined based on the category probability information, and the completion category is determined as the completion result corresponding to the missing feature; if the feature type of the missing feature is time-series, time dependency modeling is performed on the time-series data corresponding to the missing feature based on the corresponding completion guidance information to obtain the completion sequence corresponding to the missing feature, and the completion sequence is determined as the completion result corresponding to the missing feature.

[0114] Specifically, a probabilistic generative model based on a variational autoencoder can be used to complete continuous missing features. The probabilistic generative model can generate distribution parameters corresponding to the continuous missing features based on the initial completion guidance information. These distribution parameters can include a mean parameter and a variance parameter. For example, for the i-th continuous missing feature, its mean parameter can be denoted as... The variance parameter can be denoted as The imputation values ​​corresponding to continuous missing features can be derived from the mean value. variance is The imputation value is obtained by sampling from the probability distribution. This process can be expressed as: the imputation value corresponding to the i-th continuous missing feature follows a mean of variance is It follows a normal distribution.

[0115] In some implementations, for continuous features with preset value ranges or service constraints, the sampled completion values ​​can be subjected to range constraint processing. Range constraint processing may include truncation processing, resampling processing, or boundary mapping processing.

[0116] Optionally, when the missing feature is a discrete feature, the neighbor node information in the heterogeneous graph can be propagated and aggregated based on the second completion guidance information corresponding to the missing feature to obtain the category probability information corresponding to the missing feature. The completion category corresponding to the missing feature is then determined based on the category probability information, and the completion category is determined as the completion result corresponding to the missing feature.

[0117] Specifically, the second completion guidance information can be used as the neighbor attention weight or as input to generate the neighbor attention weight to control the information propagation intensity of different neighbor nodes in the heterogeneous graph. The graph neural network can propagate and aggregate information about user nodes, feature nodes, and target object nodes associated with the current discrete missing feature based on the neighbor attention weight, obtaining the category probability information corresponding to the current discrete missing feature. The category probability information can be used to characterize the probability that the discrete missing feature belongs to each candidate category. For example, for a missing occupational category feature, the category probability information can include the probabilities corresponding to candidate categories such as "corporate employee," "freelancer," "self-employed," and "other occupations." The candidate category with the highest category probability can be determined as the completion category, or the completion category can be determined from multiple candidate categories by sampling based on the category probability information. After determining the completion category, this completion category can be determined as the completion result corresponding to the discrete missing feature. For example, when the occupation category feature is missing, a heterogeneous graph can be constructed based on other basic user features, historical access behavior, target object category, and feature information of similar users. The second completion guidance information is used to control the information propagation and aggregation of relevant neighbor nodes to obtain the probability distribution of occupation category belonging to each candidate category. Then, the candidate category with the highest probability is determined as the completion result corresponding to the occupation category feature.

[0118] Optionally, when the missing feature is a temporal feature, a temporal convolutional network can be used to complete the temporal missing feature. The temporal convolutional network can include dilated causal convolutional layers. Dilated causal convolution can expand the receptive field of the convolution by setting different dilation coefficients without using future time location data, thereby capturing temporal dependencies over a longer time range. Specifically, the third completion guidance information can be used as a temporal attention weight, or used to generate temporal attention weights, to control the degree of attention the temporal model pays to information at different time locations or different time spans. Based on the temporal attention weights, the temporal model can model short-term changes, long-term trends, periodicity, and seasonality in the temporal data, generating the completed sequence corresponding to the temporal missing feature. For example, if a recent 7-day access frequency sequence is "2, missing, 3, 4, missing, 6, 5", the temporal model can predict the two missing time locations based on the existing time location data in the sequence and the third completion guidance information, obtaining the complete access frequency sequence.

[0119] It should be noted that the feature completion models used for continuous, discrete, and time-series features can be set separately, and their model parameters can be jointly updated during the training of the interaction behavior prediction model. This allows different types of feature completion models to learn completion methods that match the corresponding data structure and variation patterns.

[0120] Therefore, the embodiments of this application can fully utilize the differences in data format, association structure, and change patterns of various features to improve the rationality and accuracy of the completion results corresponding to different types of missing features. Furthermore, by jointly updating the model parameters of each feature completion model during the training process of the interactive behavior prediction model, the missing feature completion process can be adapted to the interactive behavior prediction task, thereby reducing the impact of inaccurate feature completion results on subsequent model training and interactive behavior prediction results.

[0121] S508: Generate interaction behavior completion feature information based on the completion results corresponding to each missing feature in the first interaction behavior feature information and the non-missing features in the first interaction behavior feature information.

[0122] The interaction behavior completion feature information can be the complete feature information obtained after completing the missing features in the first interaction behavior feature information. The interaction behavior completion feature information can include the completion results corresponding to each missing feature and the original feature values ​​corresponding to each non-missing feature.

[0123] Specifically, the value of each heterogeneous feature in the interaction behavior completion feature information can be determined according to the missing state corresponding to each heterogeneous feature. For heterogeneous features whose missing state indicates the presence of feature values, the original feature values ​​in the first interaction behavior feature information can be retained; for heterogeneous features whose missing state indicates the absence of feature values, the completion result corresponding to the missing feature can be written into the corresponding feature position in the first interaction behavior feature information to replace the corresponding missing marker.

[0124] For example, for the i-th heterogeneous feature, the completed feature value can be determined as follows: if the i-th heterogeneous feature is not missing, its original feature value is used; if the i-th heterogeneous feature is missing, its corresponding completed result is used. Furthermore, the completed feature information corresponding to each interaction behavior sample in the interaction behavior sample set can be combined to obtain the interaction behavior completed feature information corresponding to the interaction behavior sample set.

[0125] It should be noted that the above feature completion process can be performed only on missing features, while retaining the original feature values ​​of non-missing features. This avoids unnecessary modifications to the originally valid feature values ​​by the completion model, and ensures that the interaction behavior completion feature information includes both the original valid features and the missing feature completion results, providing input data for subsequent feature fusion, credibility filtering, and interaction behavior prediction model training.

[0126] In this embodiment, by extracting the corresponding target association pattern information from the feature association pattern information for each missing feature, the association relationship and the strength of the association between the missing feature and other heterogeneous features can be obtained. Furthermore, according to the feature type of the missing feature, the target association pattern information is converted into completion guidance information that is compatible with the corresponding feature completion process using a corresponding mapping network. Based on the completion guidance information, the completion result is generated using a completion method that matches continuous features, discrete features, or time-series features. This fully utilizes the differences in data form, association structure, and change patterns of different types of features, improving the pertinence, rationality, and accuracy of missing feature completion.

[0127] In one embodiment, see Figure 6 In S208, based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set, the initial interaction behavior prediction model is trained to obtain the target interaction behavior prediction model, which may include the following steps: S602: Determine the credibility information of the completion results corresponding to each missing feature in the interaction behavior completion feature information.

[0128] The credibility information of the completion results can be used to characterize the reliability of the completion results corresponding to the missing features. Credibility information can be represented by a credibility score, which can be within a preset numerical range, for example: the credibility score for each completion feature. ;in, This represents the confidence score of the completion result corresponding to the i-th missing feature. A higher confidence score indicates a higher level of reliability for the completion result; a lower confidence score indicates a higher level of uncertainty. The confidence information corresponding to a missing feature can be determined based on the uncertainty generated during the completion process and the data characteristics of the feature itself.

[0129] It should be noted that since the completion results of continuous features, discrete features, and time-series features have different data formats and evaluation criteria, the server can adopt the corresponding confidence determination method according to the feature type of the missing feature.

[0130] In one embodiment, in S602, determining the credibility information of the completion result corresponding to each missing feature in the interaction behavior completion feature information may include the following steps: when the feature type of the missing feature is continuous, determining the credibility information corresponding to the missing feature based on the variance information in the distribution parameters corresponding to the missing feature; when the feature type of the missing feature is discrete, determining the credibility information corresponding to the missing feature based on the maximum class probability in the class probability information corresponding to the missing feature; when the feature type of the missing feature is time-series, determining the credibility information corresponding to the missing feature based on at least one of the smoothness information and periodic consistency information of the completion sequence corresponding to the missing feature.

[0131] In one embodiment, when the missing feature is of the continuous type, the continuous feature completion model can generate the mean value based on the completion guidance information corresponding to the missing feature. and variance Then, samples are taken from the corresponding distribution to obtain the complete values ​​corresponding to the missing features.

[0132] Alternatively, the confidence score corresponding to a continuous missing feature can be determined according to the following formula: Where τ represents the temperature parameter and e represents the natural constant. The temperature parameter τ can be used to adjust the variance. Credibility score The degree of influence can be determined based on the sample set of interactive behaviors, the range of values ​​for continuous features, or the model validation results.

[0133] In another embodiment, when the missing feature is of discrete type, the category probability information generated by the discrete feature completion model for the missing feature can be obtained. It can be used to characterize the probability that a missing feature belongs to multiple candidate categories. Discrete feature completion models can propagate and aggregate neighbor node information in heterogeneous graphs to obtain the category probability distribution corresponding to the missing feature, and determine the corresponding completion category based on the category probability distribution. The maximum category probability can be category probability information. The highest probability of the missing class is the maximum class probability. A higher maximum class probability indicates that the discrete feature completion model is more likely to identify the missing feature as a candidate class, the model's judgment on the completion class is more explicit, and the reliability of the completion result is higher. A lower maximum class probability, or if the probabilities of multiple candidate classes are relatively close, indicates that the model has difficulty in clearly determining the class to which the missing feature belongs, and the uncertainty of the completion result is higher.

[0134] Alternatively, the confidence score corresponding to the discrete missing feature can be determined according to the following formula: ;in, This represents the probability distribution of the completion categories corresponding to the missing features; Indicates from probability distribution The highest probability of the selected category. For example, if a missing product preference category feature corresponds to three candidate categories, with probabilities of 0.80, 0.15, and 0.05 respectively, then... A score of 0.80 can be defined as the confidence score corresponding to the missing feature. In this case, the model clearly favors one of the candidate categories, and the corresponding completion result can be considered to have high confidence.

[0135] In another embodiment, when the missing feature is of a temporal type, a completion sequence generated by a temporal feature completion model can be obtained. Based on at least one of the smoothness information and periodic consistency information of the completion sequence, the confidence information corresponding to the missing feature can be determined. The completion sequence can be multiple feature values ​​generated by the temporal feature completion model for the missing temporal feature, arranged in chronological order. The completion sequence can be represented as: ;in, This represents the complete sequence corresponding to the i-th temporal missing feature; The completed sequence contains T real values; T represents the time length of the completed sequence. Smoothness information can be used to characterize whether the changes in the completed sequence are gradual between adjacent time points. Specifically, the first-order difference sequence of the completed sequence can be calculated first, which can be expressed as: ;in, This represents the difference between the feature value at time point (t+1) and the feature value at time point (t) in the completed sequence; This represents the feature value at time point t; This represents the feature value at the (t+1)th time point.

[0136] Furthermore, the smoothness index of the completed sequence can be calculated using the following formula: ;in, This represents the smoothness information corresponding to the completed sequence; |Δx t | represents the absolute value of the t-th first-order difference. The smaller the change between adjacent time points in the completed sequence, the better. The smaller the value, the greater the smoothness, indicating a more gradual change in the completed sequence; conversely, the more drastic the changes between adjacent time points in the completed sequence. The larger the value, the smaller the smoothness, indicating that the completed sequence may have unreasonable fluctuations.

[0137] For example, for a user's transaction frequency sequence, if the completed transaction frequency shows a significant increase or decrease between adjacent time points, it may indicate that the completed sequence does not conform to the user's usual transaction change pattern, and its smoothness can be determined to be low; if the completed transaction frequency sequence changes relatively gently, it can be determined to have high smoothness.

[0138] Optionally, periodic consistency information can be used to characterize the degree of consistency between the periodic pattern presented by the completed sequence and the expected periodic pattern of the corresponding time-series feature. Periodic consistency information can be applied to time-series features with obvious periodic characteristics, such as user repayment behavior sequences, periodic access sequences, or credit limit usage periodic pattern sequences.

[0139] Optionally, the spectral characteristics of the completed sequence can be analyzed using Fast Fourier Transform, and the dominant periodic component can be extracted from the spectral characteristics to obtain the detected dominant period length. For a sequence with an expected period length of P, the periodic consistency information can be calculated using the following formula: Where periodicity represents the periodic consistency information corresponding to the completed sequence; This indicates the dominant cycle length detected based on the completed sequence; P represents the expected cycle length corresponding to the corresponding time-series feature.

[0140] Furthermore, the smoothness information and periodic consistency information can be weighted to obtain the confidence score corresponding to the time-series missing features, which can be specifically expressed as: Here, α represents the harmonic parameter, used to adjust the weights of smoothness information and periodic consistency information in the credibility score. When the corresponding time-series feature does not have obvious periodic characteristics, α can be set to 1. In this case, the credibility score... It can be determined solely based on smoothness information, i.e.: =smoothness; When the corresponding time-series feature has obvious periodic characteristics, α can be set to a harmonic value between 0 and 1. For example, α can be set to 0.5, in which case the confidence score can be expressed as: = 0.5 × smoothness + 0.5 × periodicity; that is, smoothness information and periodic consistency information each account for half of the credibility score.

[0141] In other embodiments, the corresponding confidence information may be determined based solely on the smoothness information of the completed sequence or solely on the periodic consistency information of the completed sequence. This application does not specifically limit this.

[0142] Therefore, the embodiments of this application can fully consider the differences in data format and variation patterns of different types of features, improving the accuracy and relevance of the credibility information of the completion results. Furthermore, based on the credibility information corresponding to each completion result, a reliable basis can be provided for the subsequent dynamic filtering and weighting of interactive behavior completion feature information, thereby reducing the adverse impact of low-credibility completion results on the training and prediction results of the interactive behavior prediction model.

[0143] S604: Based on the credibility information of the completion results corresponding to each missing feature, perform feature filtering processing on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information.

[0144] Optionally, feature selection processing may include feature interaction enhancement processing and dynamic gating filtering processing. Feature interaction enhancement processing is used to establish semantic associations between the embedding information of the original features and the embedding information of the completed features, and further explore higher-order interaction relationships between different features; dynamic gating filtering processing is used to dynamically weight the interaction-enhanced features according to the credibility information of each completion result, so as to improve the contribution of high-credibility completed features to the prediction of subsequent interaction behavior, and suppress the adverse effects that low-credibility completed features may have.

[0145] In one embodiment, in S604, based on the credibility information of the completion results corresponding to each missing feature, feature filtering processing is performed on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information. This may include the following steps: embedding the interaction behavior completion feature information to obtain completion feature embedding information; using the second interaction behavior feature information as query information and the completion feature embedding information as key and value information, performing cross-attention processing on the second interaction behavior feature information and the completion feature embedding information to obtain cross-attention feature information; performing feature cross processing on the second interaction behavior feature information and the cross-attention feature information to obtain interaction enhancement feature information; determining the gating weights corresponding to each feature in the interaction enhancement feature information based on the credibility information of the completion results corresponding to each missing feature, and filtering the interaction enhancement feature information based on the gating weights to obtain the interaction behavior target feature information.

[0146] Optionally, the interaction behavior completion feature information can be embedded first to obtain completion feature embedding information. The interaction behavior completion feature information can include the completion results corresponding to each missing feature and the original feature values ​​corresponding to each non-missing feature. For continuous features, the corresponding feature values ​​can be mapped to vectors of a preset dimension through a linear projection layer; for discrete features, the corresponding categories can be mapped to vectors of a preset dimension through an embedding lookup table; for temporal features, the corresponding temporal embedding information can be generated through temporal encoding, convolution processing, or attention processing.

[0147] Therefore, different types of completion features can be transformed into a unified embedding space to obtain completion feature embedding information, which can be represented as: Subsequently, the embedding information of the original features and the embedding information of the completed features can be interactively enhanced. Specifically, the second interaction behavior feature information can be used as query information, and the embedding information of the completed features can be used as key and value information, and a cross-attention mechanism can be used to model the semantic relationship between the two. The query information can be represented as: Key information can be represented as: The value information can be represented as: .in, This indicates the characteristic information of the second interaction behavior. This indicates that the feature embedding information is complete. , and These represent the learnable query projection matrix, key projection matrix, and value projection matrix, respectively.

[0148] Furthermore, based on the above, the cross-attention feature information can be calculated using the following formula: Where CrossAttention represents cross-attention feature information, This indicates the query information generated from the second interaction behavior feature information. This represents the key information generated by the embedding information of the completed features. This represents the value information generated by the complete feature embedding information, where D represents the embedding dimension. This represents the transpose of the key information, and softmax represents the normalization function.

[0149] Through the aforementioned cross-attention processing, semantic information related to the original features can be adaptively extracted from the augmented feature embedding information based on the original feature semantics in the second interaction behavior feature information. For example, in a product recommendation scenario, information related to the current interaction behavior prediction task can be extracted from the augmented product preference features, purchase frequency features, or access sequence features based on the embedding information of user historical behavior features.

[0150] Furthermore, the cross-attention feature information and the second interaction behavior feature information can be concatenated, and the concatenated features can be input into a feature cross-network to explicitly model the high-order combination relationships between different features. The feature cross-network can adopt a deep cross-network structure. For the l-th cross-layer, the output can be represented as: ;in, This represents the initial input features after preprocessing and embedding, which may include features obtained by concatenating second interaction behavior feature information and cross attention feature information; This represents the output vector of the l-th cross layer; This represents the learnable weight matrix or weight vector corresponding to the l-th cross layer; Represents the bias term corresponding to the l-th cross layer; ⊙ represents element-wise multiplication. This represents the cross coefficient obtained based on the output of the l-th cross layer and the weight parameters.

[0151] By processing through multiple intersecting layers, higher-order combination relationships between different features can be explicitly modeled. For example, second-order or higher-order interaction relationships can be established between user attribute features, historical behavior features, target object features, and completion features to obtain interaction enhancement feature information.

[0152] After obtaining the interaction enhancement feature information, dynamic gating filtering can be applied to this information based on the credibility of the completion results corresponding to each missing feature. Specifically, for the i-th feature dimension, the corresponding gating weight can be determined according to the following formula: ;in, This represents the gating weight corresponding to the i-th feature dimension; This represents the i-th feature representation obtained after interactive enhancement of the second interactive behavior feature information and the completion feature embedding information; This represents the credibility score corresponding to the i-th completion result; This indicates that the i-th feature representation and its corresponding confidence score are concatenated. and δ represents the learnable gating weights and gating biases, respectively; δ represents the activation function Sigmoid.

[0153] Furthermore, the interaction enhancement feature information can be weighted according to the following formula: ;in, This represents the i-th feature representation after gating and filtering.

[0154] Through the aforementioned dynamic gating filtering, the impact of each completion feature on the interactive behavior prediction process can be dynamically adjusted based on the credibility information of each completion result. For completion features with high credibility, a larger gating weight can be assigned to retain their effective information; for completion features with low credibility, a smaller gating weight can be assigned to suppress the interference of unreliable completion results on the model prediction results.

[0155] Optionally, the gating weights can also be influenced by the importance information of the corresponding features. The impact of feature importance information can be used to characterize the importance of the i-th feature to the interaction behavior prediction task. For features with higher importance, a higher confidence requirement can be imposed on their completion results; when the confidence of the completion result corresponding to the important feature is low, stronger suppression can be applied to the corresponding feature. For features with lower importance, their confidence requirement can be appropriately reduced.

[0156] In one alternative implementation, feature importance information can be... Enhanced feature information through interaction And credibility score These factors are used together as input to the gating network to determine the gating weights corresponding to the i-th feature. Therefore, the reliability of the completion result and the importance of the feature itself can be combined simultaneously for more refined dynamic feature selection.

[0157] Ultimately, the filtered feature representations corresponding to each feature can be used as a basis. This generates target feature information for interactive behavior. This target feature information can include multiple feature representations after cross-attention enhancement, higher-order feature crossing, and dynamic gating filtering, and can be used as input to the initial interactive behavior prediction model to subsequently determine the interactive behavior prediction result.

[0158] Therefore, this application embodiment establishes a semantic association between the second interactive behavior feature information and the completion feature embedding information through a cross-attention mechanism, and explicitly models the high-order combination relationship between different features through a feature cross network. Then, dynamic gating filtering is performed based on the credibility information of each completion result, thereby enhancing the role of effective completion features and suppressing the adverse effects of low-credibility completion features on the interactive behavior prediction results.

[0159] S606: Input the target feature information of the interaction behavior into the initial interaction behavior prediction model to obtain the interaction behavior prediction results corresponding to each interaction behavior sample in the interaction behavior sample set.

[0160] The interactive behavior target feature information can include the target feature information corresponding to each interactive behavior sample. The target feature information corresponding to each interactive behavior sample can be used to characterize the comprehensive feature state of the corresponding target user, target object and interactive scenario.

[0161] Understandably, the initial interaction behavior prediction model can include an interaction behavior prediction network. This network can consist of multiple fully connected layers, each of which may include a linear transformation, an activation function, and Dropout regularization. The linear transformation maps the input features to a specific dimension, the activation function introduces non-linear feature representation, and Dropout regularization randomly masks some neurons during model training to reduce the likelihood of overfitting.

[0162] Specifically, the target feature information of the interaction behavior can be input into the first fully connected layer of the initial interaction behavior prediction model to map the input features to a preset hidden dimension. The processing of the first fully connected layer can be represented as follows: ;in, This represents the interactive behavior target feature information obtained after feature filtering processing; This represents the weight parameters corresponding to the first fully connected layer; This represents the bias parameters corresponding to the first fully connected layer; Represents the linear rectification activation function; This represents the hidden feature information output by the first fully connected layer.

[0163] Optionally, one or more subsequent fully connected layers can be set after the first fully connected layer. The processing procedure for the l-th fully connected layer can be expressed as follows: ;in, This represents the hidden feature information output by the (l-1)th fully connected layer; This represents the weight parameters corresponding to the l-th fully connected layer; This represents the bias parameter corresponding to the l-th fully connected layer; This represents the hidden feature information output by the l-th fully connected layer.

[0164] Optionally, Dropout regularization can be performed after each fully connected layer to randomly mask some hidden features according to a preset ratio during model training, thereby improving the model's generalization ability. After processing through one or more fully connected layers, the corresponding hidden feature information can be linearly transformed through the last output layer, and the linear transformation result can be mapped to between 0 and 1 through the Sigmoid activation function to obtain the interaction behavior prediction result corresponding to the interaction behavior sample. This process can be represented as: ;in, The value represents the predicted interaction behavior; δ represents the Sigmoid activation function. and These represent the weight parameters and bias parameters corresponding to the output layer, respectively. This represents the hidden feature information of the input and output layers.

[0165] In the embodiments of this application, the interaction behavior prediction results can be used to characterize the predicted probability that a target user will perform a target interaction behavior towards a target object. For example, in a product recommendation scenario, the interaction behavior prediction results can be used to characterize the predicted probability that a user will click on a target product; in a content recommendation scenario, the interaction behavior prediction results can be used to characterize the predicted probability that a user will click on or browse target content; in a consumer finance product recommendation scenario, the interaction behavior prediction results can be used to characterize the predicted probability that a user will click on a loan product, credit card product, or installment product.

[0166] For each interaction behavior sample in the interaction behavior sample set, the initial interaction behavior prediction model can output the corresponding interaction behavior prediction result, thereby obtaining the interaction behavior prediction result corresponding to each interaction behavior sample.

[0167] S608: Based on the interaction behavior prediction results and corresponding interaction behavior labels of each interaction behavior sample, update the parameters of the initial interaction behavior prediction model to obtain the target interaction behavior prediction model.

[0168] Interaction behavior labels can be used to characterize whether the target interaction behavior has occurred for the corresponding interaction behavior sample. For example, a first preset value can be used to indicate that the target user has performed the target interaction behavior on the target object, and a second preset value can be used to indicate that the target user has not performed the target interaction behavior on the target object. In the click behavior prediction scenario, the first preset value can be 1, and the second preset value can be 0.

[0169] In one embodiment, in S608, the initial interaction behavior prediction model is updated with parameters based on the interaction behavior prediction results and corresponding interaction behavior labels corresponding to each interaction behavior sample to obtain the target interaction behavior prediction model. This may include the following steps: determining the interaction behavior prediction loss based on the interaction behavior prediction results and corresponding interaction behavior labels corresponding to each interaction behavior sample; determining the contrastive learning loss based on the feature information corresponding to the complete feature samples, missing feature samples, and completed feature samples in the interaction behavior sample set; obtaining the attention weight information generated in the process of determining the feature association pattern information, and determining the attention regularization loss based on the attention weight information; and updating the parameters of the initial interaction behavior prediction model based on the interaction behavior prediction loss, contrastive learning loss, and attention regularization loss to obtain the target interaction behavior prediction model.

[0170] Optionally, the interaction behavior prediction loss can be determined based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample. The interaction behavior prediction loss can be used to characterize the difference between the interaction behavior prediction results output by the initial interaction behavior prediction model and the actual interaction behavior results.

[0171] Optionally, a binary cross-entropy loss function can be used to determine the interaction behavior prediction loss. The interaction behavior prediction loss can be expressed as: ;in, This represents the loss for predicting interactive behavior; N represents the number of interactive behavior samples used in this training. This represents the actual interaction behavior label corresponding to the nth interaction behavior sample; This represents the interaction behavior prediction result output by the initial interaction behavior prediction model for the nth interaction behavior sample.

[0172] Furthermore, a contrastive learning task can be introduced to enhance the model's ability to complete missing features and predict interactive behaviors. For the training samples in the interactive behavior sample set, anchor samples, completed samples, and negative samples can be constructed.

[0173] Anchor samples can be used to represent complete feature information in the full feature state. For a complete feature sample, the original complete feature information corresponding to that complete feature sample can be determined as the anchor sample.

[0174] The completed sample can be used to characterize the reconstructed feature information obtained after feature masking and feature completion processing of the complete feature sample corresponding to the anchor sample. Specifically, some feature dimensions can be randomly selected from the complete feature sample for masking to obtain missing feature samples corresponding to the complete feature sample; then, the missing features in the missing feature samples are completed by the missing feature completion module to obtain the completed sample. Since the anchor sample and the completed sample originate from the same complete feature sample, the anchor sample and the completed sample can be identified as a positive sample pair.

[0175] In one specific implementation, 30% to 60% of the feature dimensions can be randomly selected from all feature dimensions of the complete feature sample for masking to simulate different feature missing modes. The masked features are then filled in to obtain the filled sample corresponding to the complete feature sample, thereby enhancing the robustness of the model to diverse feature missing scenarios.

[0176] Negative samples can be feature information from interaction behavior samples that originate from different sources than the current anchor sample. The negative sample can include at least one of the following: complete feature information, missing feature information, or completed feature information corresponding to other interaction behavior samples within the batch. Optionally, negative samples can be determined from other interaction behavior samples that do not meet the preset positive sample matching conditions of the current anchor sample, and the current anchor sample and the negative sample can be paired to form a negative sample pair. The preset positive sample matching conditions can include at least one of the following: originating from the same interaction behavior sample, or corresponding to the same sample identifier.

[0177] Optionally, in the actual training process, an in-batch negative sampling strategy can be used to construct a set of contrastive learning samples. For a training batch with a batch size of N, for each complete feature sample, the original complete feature information corresponding to the complete feature sample can be used as an anchor sample, and some feature dimensions of the complete feature sample can be randomly masked to obtain missing feature samples; the missing feature samples are then filled in to obtain filled samples that correspond one-to-one with the anchor samples. For any anchor sample, the filled samples that originate from the same complete feature sample as the anchor sample can be determined as positive samples, and the feature information of other interactive behavior samples in the training batch that do not meet the preset positive sample matching conditions with the anchor sample can be determined as negative samples.

[0178] Understandably, in real-world applications, the number of complete feature samples may be relatively small. In such cases, relaxation strategies can be used to expand the range of samples that can serve as complete feature samples or positive samples. For example, a small number of features that are not significantly related to the target interaction behavior can be allowed to be missing, and the corresponding samples can be used as samples that meet the preset completeness condition.

[0179] Alternatively, the contrastive learning loss can be expressed as: ;in, represents the contrastive learning loss; sim represents the cosine similarity function, used to measure the semantic similarity between two feature pieces of information; This represents the anchor point feature information in a positive sample pair; The parameter τ represents the completed feature information corresponding to the completed sample; τ represents the temperature parameter; and K represents the number of negative samples corresponding to each anchor sample. The temperature parameter τ can be used to control the scaling degree of the similarity score and affect the model's attention to difficult negative samples. Among them, negative samples that have high feature similarity with anchor samples but do not meet the preset positive sample matching conditions can be considered difficult negative samples; a larger τ value can make the similarity distribution between different samples smoother.

[0180] Furthermore, attention weight information generated during the process of determining feature association patterns can be obtained, and attention regularization loss can be determined based on this information. The attention regularization loss can be used to constrain the overall distribution of attention weights, reducing the possibility of abnormal fluctuations in attention weights. The attention weight matrix generated by multi-head attention processing can be represented as: Where N represents the number of interaction behavior samples included in the training batch; H represents the number of attention heads; and F represents the number of features included in each interaction behavior sample. This represents the attention weight matrix corresponding to the nth interaction behavior sample under the hth attention head.

[0181] Alternatively, the attention regularization loss can be expressed as: ,in, This represents the attention regularization loss; Let L1 be the L1 norm of the attention weight matrix of the nth interaction sample under the hth attention head. The L1 norm can be the sum of the absolute values ​​of all elements in the corresponding attention weight matrix.

[0182] Furthermore, the model training loss can be determined based on the interaction behavior prediction loss, contrastive learning loss, and attention regularization loss. The model training loss can be expressed as: ,in, This represents the model training loss; Indicates the estimated loss from interactive behavior; Indicates the contrast learning loss; α represents the attention regularization loss; α and β represent the loss weight parameters, used to adjust the proportions of contrastive learning loss and attention regularization loss in the model training loss, respectively.

[0183] Understandably, gradient backpropagation can be performed based on the model training loss to update the model parameters in the initial interaction behavior prediction model. In one implementation, parameters in the feature embedding module, feature association pattern determination module, missing feature completion module, feature selection module, and interaction behavior prediction network can be updated simultaneously, thereby achieving end-to-end joint optimization from missing feature completion to interaction behavior prediction. During model training, an Adaptive Moment Estimation (Adam) optimizer can be used to update the model parameters, and a cosine annealing strategy can be used to dynamically adjust the learning rate. For example, the initial learning rate can be set to 0.001, and the minimum learning rate can be set to 1×10⁻⁶. －6 The training cycle is set to 20 epochs. Forward computation, model training loss determination, and gradient backpropagation can be repeatedly performed according to the preset training cycle until the preset training termination conditions are met. The preset training termination conditions may include at least one of the following: the model training loss reaches convergence, the training cycle reaches a preset number of epochs, and the model's predicted performance on the validation sample set meets a preset performance requirement. After meeting the preset training termination conditions, the initial interaction behavior prediction model with updated parameters can be designated as the target interaction behavior prediction model.

[0184] Therefore, by jointly utilizing the interaction behavior prediction loss, contrastive learning loss, and attention regularization loss to update model parameters, this embodiment of the application can improve the accuracy of the interaction behavior prediction results, enhance the model's ability to complete missing features, and improve the model's ability to extract key feature correlations, thereby achieving end-to-end joint optimization of the feature completion process and the interaction behavior prediction process.

[0185] In this embodiment, by first determining the credibility information of the completion results corresponding to each missing feature, and then combining the credibility information to enhance the feature interaction of the second interactive behavior feature information and the interactive behavior completion feature information and perform dynamic filtering, it is possible to fully utilize the effective information carried by the completion features while suppressing the interference of low-credibility completion results on the subsequent prediction process, thus obtaining more reliable interactive behavior target feature information. Furthermore, by inputting the interactive behavior target feature information into the initial interactive behavior prediction model, and jointly updating the model parameters by combining the interactive behavior prediction loss, contrastive learning loss, and attention regularization loss, it is possible to simultaneously improve the model's interactive behavior prediction ability, missing feature restoration ability, and key feature association extraction ability, thereby achieving end-to-end collaborative optimization of feature completion, feature filtering, and interactive behavior prediction, and improving the prediction accuracy, stability, and generalization ability of the target interactive behavior prediction model in feature missing scenarios.

[0186] The following combination Figure 7 This application is described below. Figure 7 This is a schematic diagram illustrating the overall process of training an interactive behavior prediction model, provided as an exemplary embodiment of this application. Figure 7 First, the raw feature input can be obtained, which may include multiple heterogeneous features corresponding to each interaction behavior sample in the interaction behavior sample set. Then, the raw feature input can be preprocessed in two spaces to generate raw value representations and embedded representations. The raw value representation corresponds to the raw value space and can retain the original feature values ​​and missing states of each heterogeneous feature; the embedded representation corresponds to the embedded space and can be used for semantic interactions between different heterogeneous features within the model.

[0187] In the embedding space, association pattern discovery can be performed based on the embedding representation to determine the association relationships and strengths between different heterogeneous features, and further learn the mapping relationship from association patterns to completion guidance information. Completion guidance information can be used to guide the completion of missing features in the original value space. In the original value space, multimodal completion of missing features can be performed based on completion guidance information. For continuous, discrete, and time-series features, feature completion methods adapted to their data form and variation patterns can be adopted to generate corresponding completion values, completion categories, or completion sequences, while retaining the original feature values ​​of the features that are not missing, thereby obtaining interactive behavior completion feature information.

[0188] Furthermore, the quality of the completion results corresponding to each missing feature can be evaluated to obtain the corresponding credibility information, and the interaction behavior completion feature information can be embedded and transformed. Subsequently, feature interaction enhancement can be performed on the embedded representations of the original features and the embedded representations of the completion features to establish semantic associations between the two types of features and mine higher-order feature combination relationships; then, dynamic gating filtering is performed based on the credibility information of each completion result to improve the retention of high-credibility completion features and suppress the impact of low-credibility completion features on the subsequent prediction process.

[0189] Finally, the target feature information, after feature interaction enhancement and dynamic gating filtering, can be input into the interaction behavior prediction network to obtain the interaction behavior prediction result. During model training, the model training loss can be determined by combining the interaction behavior prediction result with the corresponding interaction behavior label, and the model parameters of each processing module and the interaction behavior prediction network can be jointly updated based on the model training loss to obtain the target interaction behavior prediction model.

[0190] thus, Figure 7 The process shown involves dual-space preprocessing, association pattern discovery, multimodal feature completion, completion quality assessment, feature interaction enhancement, and dynamic gating filtering. This process achieves collaborative optimization between the missing feature completion process and the interaction behavior prediction process, thereby improving the prediction accuracy, stability, and generalization ability of the interaction behavior prediction model in scenarios with missing features.

[0191] Furthermore, after obtaining the target interaction behavior prediction model, interaction behavior data to be predicted can be acquired. The interaction behavior data to be predicted includes multiple heterogeneous features, including missing features. The interaction behavior data to be predicted is used to characterize the interaction scenario data between the target user and the target object. Then, the interaction behavior data to be predicted is input into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be predicted. The interaction behavior prediction result is used to characterize the predicted probability that the target user will perform the target interaction behavior towards the target object.

[0192] The interactive behavior data to be estimated may include multiple heterogeneous features such as the current user's user attribute characteristics, historical behavior characteristics, target product attribute characteristics, and current recommendation scenario characteristics. For example, the user attribute characteristics may include occupation category, the historical behavior characteristics may include login frequency in the past 30 days, number of product page visits in the past 7 days, and historical click ratio, the target product attribute characteristics may include product category, product price range, and product brand, and the current recommendation scenario characteristics may include recommendation time, terminal type, and access channel.

[0193] In practical applications, the interactive behavior data to be predicted may contain missing features. For example, if the current user is a newly registered user, features such as their historical click ratio and historical purchase frequency may be missing; or due to data synchronization delays, the number of times the current user has visited product pages in the past 7 days may not have been updated, which may also lead to the missing of corresponding features. In this case, the interactive behavior data to be predicted, including the missing features, can be input into the target interactive behavior prediction model.

[0194] It is understandable that the target interaction behavior prediction model can perform the processing procedures corresponding to the training phase on the interaction behavior data to be predicted. Specifically, the target interaction behavior prediction model can perform dual-space feature processing on the interaction behavior data to be predicted, obtaining first interaction behavior feature information corresponding to the original value space and second interaction behavior feature information corresponding to the embedding space; determine the feature association pattern information between each heterogeneous feature based on the second interaction behavior feature information; perform feature completion processing on the missing features in the first interaction behavior feature information based on the feature association pattern information, obtaining completed interaction behavior feature information; and perform feature filtering processing and interaction behavior prediction processing based on the completed interaction behavior feature information, outputting the interaction behavior prediction result. The interaction behavior prediction result can be used to characterize the predicted probability of the current user performing a click behavior on the target product. For example, if the interaction behavior prediction result output by the target interaction behavior prediction model is 0.82, it can indicate that the predicted probability of the current user clicking on the target product is high; if the output interaction behavior prediction result is 0.15, it can indicate that the predicted probability of the current user clicking on the target product is low. The server can determine the ranking position of the target product in the recommendation list based on the interaction behavior prediction results, or determine whether to display the target product to the current user based on the interaction behavior prediction results.

[0195] Therefore, in the model application stage, the interactive behavior data to be predicted, including missing features, can be directly input into the target interactive behavior prediction model. The target interactive behavior prediction model then processes the missing features and outputs the interactive behavior prediction result. By matching the processing flow in the inference stage with the missing feature completion and feature selection process in the training stage, the impact of missing features in the interactive behavior data to be predicted on the prediction result can be reduced, improving the accuracy and stability of the interactive behavior prediction result, and providing a more reliable basis for downstream processing such as target object recommendation, ranking, or display.

[0196] It should be understood that although the steps in the flowcharts of the above embodiments 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 above embodiments 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.

[0197] Based on the above-mentioned interactive behavior prediction model training method, such as Figure 8 As shown in the embodiments of this application, an interaction behavior prediction model training device 800 is also provided for implementing the interaction behavior prediction model training method described above. The interaction behavior prediction model training device 800 includes: The first acquisition module 810 is used to acquire first interaction behavior feature information and second interaction behavior feature information corresponding to the interaction behavior sample set. The interaction behavior sample set includes multiple heterogeneous features. The first interaction behavior feature information corresponds to the original value space, and the second interaction behavior feature information corresponds to the embedding space. The determination module 820 is used to determine the feature association pattern information between heterogeneous features based on the second interaction behavior feature information; Processing module 830 is used to complete the missing features in the first interactive behavior feature information based on feature association pattern information to obtain interactive behavior completion feature information. Training module 840 is used to train the initial interaction behavior prediction model based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set, so as to obtain the target interaction behavior prediction model.

[0198] In one possible implementation, the first acquisition module 810 is specifically used for: Obtain the interaction behavior sample set and determine the missing state corresponding to each heterogeneous feature of each interaction behavior sample in the interaction behavior sample set. For each interaction behavior sample, the first interaction behavior feature sub-information corresponding to the interaction behavior sample is generated based on the feature values ​​and missing states of each heterogeneous feature in the interaction behavior sample. For each interaction behavior sample, based on the feature type and missing state of each heterogeneous feature in the interaction behavior sample, the heterogeneous features are embedded to generate the second interaction behavior feature sub-information corresponding to the interaction behavior sample. Based on the first interaction behavior feature sub-information corresponding to each interaction behavior sample, the first interaction behavior feature information corresponding to the interaction behavior sample set is determined, wherein the first interaction behavior feature sub-information includes the original feature value corresponding to the non-missing heterogeneous feature and the missing label corresponding to the missing heterogeneous feature. Based on the second interaction behavior feature sub-information corresponding to each interaction behavior sample, the second interaction behavior feature information corresponding to the interaction behavior sample set is determined. The second interaction behavior feature sub-information includes feature embedding information corresponding to heterogeneous features that are not missing and null value embedding information corresponding to missing heterogeneous features.

[0199] In one possible implementation, module 820 is specifically used for: Multi-head attention processing is performed on the feature information of the second interaction behavior to obtain attention weight information; Based on attention weight information, determine the correlation strength information between any two heterogeneous features among multiple heterogeneous features; Based on the correlation strength information between any two heterogeneous features, feature association pattern information is generated.

[0200] In one possible implementation, the processing module 830 is specifically used for: For each missing feature in the first interaction behavior feature information, the target association pattern information corresponding to the missing feature is obtained from the feature association pattern information; Based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature. Based on the completion guidance information corresponding to the missing features, the missing features are completed to obtain the completion results corresponding to the missing features. Based on the completion results corresponding to each missing feature in the first interaction behavior feature information and the non-missing features in the first interaction behavior feature information, interaction behavior completion feature information is generated.

[0201] In one possible implementation, the processing module 830 is specifically used for: When the missing feature is of continuous type, the first mapping network is used to map the target association pattern information corresponding to the missing feature to obtain the first completion guidance information for determining the distribution parameter corresponding to the missing feature. When the missing feature is of discrete type, a second mapping network is used to map the target association pattern information corresponding to the missing feature to obtain the second completion guidance information used to guide the information propagation in the heterogeneous graph. When the missing feature is of time-series type, a third mapping network is used to map the target association pattern information corresponding to the missing feature to obtain third completion guidance information for guiding time-dependent modeling.

[0202] In one possible implementation, the processing module 830 is specifically used for: When the missing feature is of continuous type, the distribution parameters corresponding to the missing feature are generated based on the corresponding completion guidance information, and the completion value corresponding to the missing feature is obtained by sampling from the probability distribution corresponding to the distribution parameters. The completion value is then determined as the completion result corresponding to the missing feature. When the missing feature is discrete, the neighbor node information in the heterogeneous graph is propagated and aggregated based on the corresponding completion guidance information to obtain the category probability information corresponding to the missing feature. The completion category corresponding to the missing feature is determined based on the category probability information, and the completion category is determined as the completion result corresponding to the missing feature. When the missing feature is of time-series type, time-dependency modeling is performed on the time-series data corresponding to the missing feature based on the corresponding completion guidance information to obtain the completion sequence corresponding to the missing feature, and the completion sequence is determined as the completion result corresponding to the missing feature.

[0203] In one possible implementation, the training module 840 is specifically used for: Determine the credibility information of the completion results corresponding to each missing feature in the feature information of interactive behavior completion; Based on the credibility information of the completion results corresponding to each missing feature, feature filtering processing is performed on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information. Input the target feature information of the interaction behavior into the initial interaction behavior prediction model to obtain the interaction behavior prediction results corresponding to each interaction behavior sample in the interaction behavior sample set. Based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

[0204] In one possible implementation, the training module 840 is specifically used for: When the missing feature is of continuous type, the confidence information corresponding to the missing feature is determined based on the variance information in the distribution parameters corresponding to the missing feature. When the missing feature is discrete, the confidence information corresponding to the missing feature is determined based on the maximum class probability in the class probability information corresponding to the missing feature. When the missing feature is of time-series type, the confidence information corresponding to the missing feature is determined based on at least one of the smoothness information and periodic consistency information of the completed sequence corresponding to the missing feature.

[0205] In one possible implementation, the training module 840 is specifically used for: The interaction behavior completion feature information is embedded to obtain the completion feature embedding information; Using the second interaction behavior feature information as query information and the completion feature embedding information as key and value information, cross-attention processing is performed on the second interaction behavior feature information and the completion feature embedding information to obtain cross-attention feature information. The second interaction behavior feature information and the cross attention feature information are subjected to feature cross processing to obtain interaction enhancement feature information; Based on the credibility information of the completion results corresponding to each missing feature, the gating weights corresponding to each feature in the interaction enhancement feature information are determined, and the interaction enhancement feature information is filtered based on the gating weights to obtain the interaction behavior target feature information.

[0206] In one possible implementation, the training module 840 is specifically used for: Based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the interaction behavior prediction loss is determined. Based on the feature information corresponding to complete feature samples, missing feature samples and complete feature samples in the interaction behavior sample set, the contrastive learning loss is determined. Obtain attention weight information generated during the process of determining feature association pattern information, and determine attention regularization loss based on attention weight information; Based on the interaction behavior prediction loss, contrastive learning loss, and attention regularization loss, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

[0207] The division of modules in the above-described interactive behavior prediction model training device is for illustrative purposes only. In other embodiments, the interactive behavior prediction model training device can be divided into different modules as needed to complete all or part of the functions of the interactive behavior prediction model training device. The implementation of each module in the interactive behavior prediction model training device provided in this application embodiment can be in the form of a computer program. This computer program can run on a terminal or server. The program modules constituted by this computer program can be stored in the memory of the terminal or server. When the computer program is executed by a processor, it implements all or part of the steps of the interactive behavior prediction model training method described in this application embodiment.

[0208] Based on the above-mentioned interactive behavior prediction method, such as Figure 9 As shown in the illustration, this application also provides an interaction behavior prediction device 900 for implementing the interaction behavior prediction method described above. The interaction behavior prediction device 900 includes: The second acquisition module 910 is used to acquire the interaction behavior data to be estimated. The interaction behavior data to be estimated includes multiple heterogeneous features, including missing features. The interaction behavior data to be estimated is used to characterize the interaction scenario data between the target user and the target object. The inference module 920 is used to input the interaction behavior data to be predicted into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be predicted. The interaction behavior prediction result is used to characterize the predicted probability of the target user performing the target interaction behavior on the target object. The target interaction behavior prediction model is generated by the interaction behavior prediction model training method provided in this application.

[0209] The division of modules in the above-described interactive behavior prediction device is for illustrative purposes only. In other embodiments, the interactive behavior prediction device can be divided into different modules as needed to complete all or part of the functions of the interactive behavior prediction device. The implementation of each module in the interactive behavior prediction device provided in this application embodiment can be in the form of a computer program. This computer program can run on a terminal or server. The program modules constituted by this computer program can be stored in the memory of the terminal or server. When the computer program is executed by a processor, it implements all or part of the steps of the interactive behavior prediction method described in this application embodiment.

[0210] This application also provides an electronic device, which can be a server, and its internal structure diagram can be as follows: Figure 10 As shown, this electronic device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O 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, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. The processor executes computer programs to implement an interactive behavior prediction model training method or an interactive behavior prediction method.

[0211] Those skilled in the art will understand that Figure 10 The structures shown are merely block diagrams of some structures related to the embodiments of this application and do not constitute a limitation on the electronic devices to which the embodiments of this application are applied. Specific electronic devices may include more or fewer components than those shown in the figures, or combine certain components, or have different component arrangements.

[0212] In one possible implementation, a computer storage medium is provided that stores instructions, which, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the above embodiments. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium.

[0213] In one possible implementation, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0214] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer storage medium or transmitted through the computer storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The aforementioned available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0215] It should be noted that the information (including but not limited to the interactive behavior sample set, the first interactive behavior feature information, and the second interactive behavior feature information), data (including but not limited to data used for analysis, stored data, and displayed data), and signals involved in the embodiments of this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the interactive behavior sample set involved in this application was obtained under full authorization.

[0216] 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. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0217] The above-described embodiments are merely preferred embodiments of the present application and are not intended to limit the scope of the present application. Various modifications and improvements made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application should fall within the protection scope defined by the claims.

[0218] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for training an interactive behavior prediction model, comprising: Obtain first interaction behavior feature information and second interaction behavior feature information corresponding to the interaction behavior sample set. The interaction behavior sample set includes multiple heterogeneous features. The first interaction behavior feature information corresponds to the original value space, and the second interaction behavior feature information corresponds to the embedding space. Based on the second interaction behavior feature information, the feature association pattern information between the heterogeneous features is determined; Based on the feature association pattern information, the missing features in the first interactive behavior feature information are completed to obtain interactive behavior completion feature information. Based on the completed feature information of the interaction behavior and the interaction behavior labels corresponding to the interaction behavior sample set, the initial interaction behavior prediction model is trained to obtain the target interaction behavior prediction model.

2. The method as described in claim 1, wherein obtaining the first interaction behavior feature information and the second interaction behavior feature information corresponding to the interaction behavior sample set includes: Obtain an interactive behavior sample set and determine the missing state corresponding to each heterogeneous feature of each interactive behavior sample in the interactive behavior sample set. For each interaction behavior sample, based on the feature values ​​and missing states of each heterogeneous feature in the interaction behavior sample, a first interaction behavior feature sub-information corresponding to the interaction behavior sample is generated; For each interaction behavior sample, based on the feature type and missing state of each heterogeneous feature in the interaction behavior sample, the heterogeneous features are embedded to generate the second interaction behavior feature sub-information corresponding to the interaction behavior sample. Based on the first interaction behavior feature sub-information corresponding to each interaction behavior sample, the first interaction behavior feature information corresponding to the interaction behavior sample set is determined, wherein the first interaction behavior feature sub-information includes the original feature value corresponding to the non-missing heterogeneous feature and the missing label corresponding to the missing heterogeneous feature. Based on the second interaction behavior feature sub-information corresponding to each interaction behavior sample, the second interaction behavior feature information corresponding to the interaction behavior sample set is determined, wherein the second interaction behavior feature sub-information includes feature embedding information corresponding to non-missing heterogeneous features and null value embedding information corresponding to missing heterogeneous features.

3. The method as described in claim 1, wherein determining the feature association pattern information between the heterogeneous features based on the second interaction behavior feature information includes: Multi-head attention processing is performed on the second interaction behavior feature information to obtain attention weight information; Based on the attention weight information, determine the correlation strength information between any two heterogeneous features among the plurality of heterogeneous features; Based on the correlation strength information between any two heterogeneous features, feature association pattern information is generated.

4. The method as described in claim 1, wherein the step of performing completion processing on the missing features in the first interaction behavior feature information based on the feature association pattern information to obtain interaction behavior completion feature information includes: For each missing feature in the first interactive behavior feature information, the target association pattern information corresponding to the missing feature is obtained from the feature association pattern information; Based on the feature type of the missing feature, the target association pattern information corresponding to the missing feature is mapped to obtain the completion guidance information corresponding to the missing feature; Based on the completion guidance information corresponding to the missing feature, the missing feature is completed to obtain the completion result corresponding to the missing feature; Based on the completion results corresponding to each missing feature in the first interaction behavior feature information and the non-missing features in the first interaction behavior feature information, interaction behavior completion feature information is generated.

5. The method as described in claim 4, wherein mapping the target association pattern information corresponding to the missing feature based on the feature type of the missing feature to obtain the completion guidance information corresponding to the missing feature includes: When the feature type of the missing feature is continuous, a first mapping network is used to map the target association pattern information corresponding to the missing feature to obtain first completion guidance information for determining the distribution parameter corresponding to the missing feature. When the feature type of the missing feature is discrete, a second mapping network is used to map the target association pattern information corresponding to the missing feature to obtain second completion guidance information for guiding the information propagation in heterogeneous graphs. When the missing feature is of time-series type, a third mapping network is used to map the target association pattern information corresponding to the missing feature to obtain third completion guidance information for guiding time-dependent modeling.

6. The method as described in claim 4, wherein the step of performing completion processing on the missing feature based on the completion guidance information corresponding to the missing feature to obtain the completion result corresponding to the missing feature includes: When the missing feature is of continuous type, the distribution parameters corresponding to the missing feature are generated based on the corresponding completion guidance information, and the completion value corresponding to the missing feature is obtained by sampling from the probability distribution corresponding to the distribution parameters, and the completion value is determined as the completion result corresponding to the missing feature. When the feature type of the missing feature is discrete, the neighbor node information in the heterogeneous graph is propagated and aggregated based on the corresponding completion guidance information to obtain the category probability information corresponding to the missing feature. The completion category corresponding to the missing feature is determined based on the category probability information, and the completion category is determined as the completion result corresponding to the missing feature. When the missing feature is of time-series type, time-dependency modeling is performed on the time-series data corresponding to the missing feature based on the corresponding completion guidance information to obtain the completion sequence corresponding to the missing feature, and the completion sequence is determined as the completion result corresponding to the missing feature.

7. The method as described in claim 1, wherein training the initial interaction behavior prediction model based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set to obtain the target interaction behavior prediction model includes: Determine the credibility information of the completion results corresponding to each missing feature in the completion feature information of the interaction behavior; Based on the credibility information of the completion results corresponding to each missing feature, feature filtering processing is performed on the second interaction behavior feature information and the interaction behavior completion feature information to obtain the interaction behavior target feature information. The target feature information of the interaction behavior is input into the initial interaction behavior prediction model to obtain the interaction behavior prediction result corresponding to each interaction behavior sample in the interaction behavior sample set. Based on the interaction behavior prediction results and corresponding interaction behavior labels corresponding to each interaction behavior sample, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

8. The method of claim 7, wherein determining the credibility information of the completion result corresponding to each missing feature in the interaction behavior completion feature information includes: When the feature type of the missing feature is continuous, the confidence information corresponding to the missing feature is determined based on the variance information in the distribution parameters corresponding to the missing feature. When the feature type of the missing feature is discrete, the credibility information corresponding to the missing feature is determined based on the maximum class probability in the class probability information corresponding to the missing feature. When the feature type of the missing feature is time-series, the credibility information corresponding to the missing feature is determined based on at least one of the smoothness information and periodic consistency information of the completed sequence corresponding to the missing feature.

9. The method as described in claim 7, wherein the step of performing feature filtering processing on the second interaction behavior feature information and the interaction behavior completion feature information based on the credibility information of the completion results corresponding to each missing feature to obtain interaction behavior target feature information includes: The interaction behavior completion feature information is embedded to obtain completion feature embedding information; Using the second interactive behavior feature information as query information and the completion feature embedding information as key information and value information, cross-attention processing is performed on the second interactive behavior feature information and the completion feature embedding information to obtain cross-attention feature information. The second interactive behavior feature information and the cross attention feature information are subjected to feature cross processing to obtain interactive enhancement feature information; Based on the credibility information of the completion results corresponding to each missing feature, the gating weights corresponding to each feature in the interaction enhancement feature information are determined, and the interaction enhancement feature information is filtered based on the gating weights to obtain the interaction behavior target feature information.

10. The method of claim 7, wherein updating the parameters of the initial interaction behavior prediction model based on the interaction behavior prediction results and corresponding interaction behavior labels corresponding to each interaction behavior sample to obtain the target interaction behavior prediction model includes: Based on the interaction behavior prediction results and corresponding interaction behavior labels for each interaction behavior sample, the interaction behavior prediction loss is determined. Based on the feature information corresponding to the complete feature samples, missing feature samples and complete feature samples in the interaction behavior sample set, the contrastive learning loss is determined. Obtain attention weight information generated during the process of determining the feature association pattern information, and determine attention regularization loss based on the attention weight information; Based on the interaction behavior prediction loss, the contrastive learning loss, and the attention regularization loss, the parameters of the initial interaction behavior prediction model are updated to obtain the target interaction behavior prediction model.

11. An interactive behavior prediction method, comprising: Acquire interactive behavior data to be estimated, the interactive behavior data to be estimated includes multiple heterogeneous features, the multiple heterogeneous features include missing features, the interactive behavior data to be estimated is used to characterize the interaction scenario data between the target user and the target object; The interaction behavior data to be predicted is input into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be predicted. The interaction behavior prediction result is used to characterize the predicted probability that the target user will perform the target interaction behavior for the target object. The target interaction behavior prediction model is generated by the interaction behavior prediction model training method as described in any one of claims 1 to 10.

12. A training device for an interactive behavior prediction model, comprising: The first acquisition module is used to acquire first interaction behavior feature information and second interaction behavior feature information corresponding to the interaction behavior sample set. The interaction behavior sample set includes multiple heterogeneous features. The first interaction behavior feature information corresponds to the original value space, and the second interaction behavior feature information corresponds to the embedding space. The determining module is used to determine the feature association pattern information between the heterogeneous features based on the second interaction behavior feature information; The processing module is used to complete the missing features in the first interactive behavior feature information based on the feature association pattern information to obtain interactive behavior completion feature information. The training module is used to train the initial interaction behavior prediction model based on the interaction behavior completion feature information and the interaction behavior labels corresponding to the interaction behavior sample set, so as to obtain the target interaction behavior prediction model.

13. An interactive behavior prediction device, comprising: The second acquisition module is used to acquire interactive behavior data to be estimated. The interactive behavior data to be estimated includes multiple heterogeneous features, including missing features. The interactive behavior data to be estimated is used to characterize the interaction scenario data between the target user and the target object. The inference module is used to input the interaction behavior data to be estimated into the target interaction behavior prediction model to obtain the interaction behavior prediction result output by the target interaction behavior prediction model for the interaction behavior data to be estimated. The interaction behavior prediction result is used to characterize the predicted probability that the target user will perform the target interaction behavior for the target object. The target interaction behavior prediction model is generated by the interaction behavior prediction model training method as described in any one of claims 1 to 10.

14. An electronic device comprising: Processor and memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code stored in the memory to perform the method as described in any one of claims 1-11.

15. A computer storage medium storing a plurality of instructions adapted for loading by a processor and performing the method steps of any one of claims 1-11.

16. A computer program product comprising instructions that, when run on a computer or processor, cause the computer or processor to perform the steps of the method as claimed in any one of claims 1-11.

Images