Data back transmission method and device based on double-model joint estimation, equipment and medium

Abstract

The present disclosure provides a data backhaul method and device based on double model joint estimation, an electronic device and a medium, relating to the technical field of Internet. The present disclosure is as follows: obtaining static attribute data and dynamic behavior data of a user terminal device within an initial time window due to a target application activation event; extracting features of the static attribute data and the dynamic behavior data to obtain feature data; inputting the feature data into a pre-trained first estimation model and a second estimation model respectively to obtain a revenue prediction value and a behavior prediction value, wherein the revenue prediction value represents the cumulative revenue of a target user within a future preset period, and the behavior prediction value represents the number of occurrences of a specific interaction behavior of the target user within the future preset period; matching the revenue prediction value, the behavior prediction value and a pre-configured joint backhaul condition to obtain a matching result; and in response to the matching result indicating that the joint backhaul condition is met, the activation event is regarded as a target conversion sample and is backhauled to an advertisement delivery platform.

Description

Technical Field

[0001] This disclosure relates to the field of Internet technology, particularly to the fields of advertising and machine learning, and specifically to a data backhaul method and apparatus, electronic device and computer-readable storage medium based on dual-model joint prediction. Background Technology

[0002] In mobile internet advertising, key behavior-based user acquisition (also known as target conversion bidding or deep conversion optimization) has become the mainstream model. Advertisers send specific deep conversion behaviors of activated users (such as paid subscriptions, consumption of key content, and incentivized video views) back to the advertising platform (media). The media then uses this feedback data to train models and find new users with similar behavioral potential.

[0003] Currently, advertisers generally adopt a fixed-rule feedback strategy based on actual user behavior. For example, a user is only sent back to the media as a positive sample when the total actual revenue generated after activation reaches 3.9 yuan and the actual number of impressions of the incentive video ad reaches 15. However, this has the following obvious drawbacks: High latency in data transmission means users need a considerable amount of time to accumulate to the preset threshold, resulting in severely delayed data transmission signals that cannot promptly guide the media model's real-time learning and optimization. Low learning efficiency means the correlation between data transmission signals and long-term user value is weak, and due to latency, the media model struggles to quickly identify high-value users during the cold start phase, limiting the improvement of advertising spending capacity. Furthermore, it cannot predict future value; existing methods can only make judgments based on past behavior and cannot predict in advance whether a newly activated user will become a high-value, heavy user in the future, thus missing the opportunity to seize high-quality traffic. Summary of the Invention

[0004] This disclosure provides a data backhaul method and apparatus, electronic device, and computer-readable storage medium based on dual-model joint prediction.

[0005] According to the first aspect, a data feedback method based on dual-model joint prediction is provided. This method includes: acquiring static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to a target application activation event; extracting features from the static attribute data and dynamic behavior data to obtain feature data of the target user; inputting the feature data into a pre-trained first prediction model and a second prediction model to obtain a revenue prediction value output by the first prediction model and a behavior prediction value output by the second prediction model, wherein the revenue prediction value represents the cumulative revenue of the target user within a preset future period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs within the preset future period; matching the revenue prediction value and the behavior prediction value with pre-configured joint feedback conditions to obtain a matching result; and responding to the matching result indicating that the joint feedback conditions are met, feeding the activation event back as a target conversion sample to the advertising platform for model optimization on the advertising platform.

[0006] According to the second aspect, a data feedback device based on dual-model joint prediction is provided. The device includes: an acquisition unit configured to acquire static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to a target application activation event; an extraction unit configured to extract features from the static attribute data and dynamic behavior data to obtain feature data of the target user; an input unit configured to input the feature data into a pre-trained first prediction model and a second prediction model respectively to obtain a revenue prediction value output by the first prediction model and a behavior prediction value output by the second prediction model, wherein the revenue prediction value represents the cumulative revenue of the target user within a future preset period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs within the future preset period; a matching unit configured to match the revenue prediction value and the behavior prediction value with pre-configured joint feedback conditions to obtain a matching result; and a feedback unit configured to, in response to the matching result indicating that the joint feedback conditions are met, feed the activation event back as a target conversion sample to an advertising platform for model optimization on the advertising platform.

[0007] According to a third aspect, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method as described in any implementation of the first aspect.

[0008] According to a fourth aspect, a non-transitory computer-readable storage medium is provided that stores computer instructions for causing a computer to perform the method described in any implementation of the first aspect.

[0009] The embodiments of this disclosure provide a data feedback method and apparatus based on dual-model joint prediction. First, static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to a target application activation event are acquired. Second, features of the static attribute data and dynamic behavior data are extracted to obtain feature data of the target user. Third, the feature data are input into a pre-trained first prediction model and a second prediction model to obtain a revenue prediction value output by the first prediction model and a behavior prediction value output by the second prediction model. The revenue prediction value represents the cumulative revenue of the target user within a preset future period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs within the preset future period. Then, the revenue prediction value and the behavior prediction value are matched with pre-configured joint feedback conditions to obtain a matching result. Finally, in response to the matching result indicating that the joint feedback conditions are met, the activation event is fed back as a target conversion sample to the advertising platform for model optimization. Therefore, through model prediction, high-value activation events can be quickly identified and transmitted back based on user behavior and profiles after user activation, greatly reducing transmission delay; more accurate and timely high-value information can be transmitted back to the media, which can help the advertising platform's model learn and expand the delivery volume more quickly; in the cold start stage where users have no historical behavior or sparse behavior, the model can make effective predictions based on limited features, overcoming the shortcomings of traditional methods that cannot effectively transmit back information in the cold start stage.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure.

[0013] Figure 1 This is a flowchart of an embodiment of the data backhaul method based on dual-model joint prediction according to this disclosure; Figure 2 This is a flowchart of another embodiment of the data backhaul method based on dual-model joint prediction according to this disclosure; Figure 3This is a schematic diagram of a structure of an embodiment of a data backhaul device based on dual-model joint prediction according to the present disclosure; Figure 4 This is a block diagram of an electronic device used to implement the data backhaul method based on dual-model joint prediction in the embodiments of this disclosure. Detailed Implementation

[0014] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0015] The technical solutions of this disclosure are illustrated below through specific embodiments. It should be understood that one or more steps mentioned in this disclosure do not preclude the existence of other methods and steps before or after the combined steps, or that other methods and steps may be inserted between these explicitly mentioned steps. It should also be understood that these examples are for illustrative purposes only and are not intended to limit the scope of this disclosure. Unless otherwise stated, the numbering of each method step is only for the purpose of identifying each method step, and not for limiting the order of each method or limiting the scope of implementation of this disclosure. Changes or adjustments to their relative relationships, without substantial changes to the technical content, can also be considered as within the scope of implementation of this disclosure.

[0016] The raw materials and instruments used in the examples are not subject to any specific restrictions on their source; they can be purchased from the market or prepared according to conventional methods known to those skilled in the art.

[0017] To address the shortcomings of traditional technologies, this disclosure proposes a data feedback method based on dual-model joint prediction. This method abandons the traditional approach of waiting for the accumulation of real user behavior, and instead uses a machine learning model to predict the user's future long-term value and the number of times to incentivize video ad exposures shortly after the user is activated, and provides real-time feedback on high-value users based on the joint prediction results. Figure 1 The flowchart 100 illustrates an embodiment of the data backhaul method based on dual-model joint prediction according to this disclosure, which includes the following steps: Step 101: Obtain static attribute data and dynamic behavior data of the user terminal device within a preset initial time window due to the target application activation event.

[0018] In this embodiment, the initial time window refers to the time window for collecting data after the user activates the device (e.g., 15 minutes).

[0019] In this embodiment, data feedback, also known as user feedback, refers to the process by which advertisers (application providers) send user data related to specific deep conversion behaviors (such as payment, generating high advertising revenue, etc.) to the advertising platform (media) through API (Application Programming Interface) interfaces, under the internet advertising key behavior user acquisition (deep conversion optimization) model. After receiving this positive sample, the media uses it to optimize its underlying advertising distribution model.

[0020] In this embodiment, the activation event is an event sent by the user to the target application, such as restarting the target application. Before obtaining static attribute data and dynamic behavior data, the data feedback method based on the dual-model joint prediction is applied to the execution entity to monitor the user's activation event to the target application.

[0021] In this embodiment, after detecting an activation event of the target application on the user terminal device, the system determines a preset initial time window starting from the trigger time of the activation event, and synchronously / asynchronously collects terminal-side data within this time window. Static attribute data includes, but is not limited to, information that does not change with short-term operations, such as region, device model, operating system version, channel, application version, network standard, resolution, language and time zone, and device identifier (after de-identification / hashing). Dynamic behavior data includes, but is not limited to, event streams generated by changes in user interaction or running status, such as foreground / background switching, page / function click sequences, dwell time, startup time, crash / lag logs, permission pop-up responses, network request latency, and sensor / location status changes. Furthermore, dynamic behavior data may also include: the number of times the target application is launched, the number of articles read, the number of videos watched, advertising revenue, and task rewards. The static attribute data and dynamic behavior data are aligned and packaged according to a unified timestamp to form user terminal behavior profile data corresponding to the initial time window for subsequent analysis and processing.

[0022] In this embodiment, the execution entity running on the data backhaul method based on dual-model joint prediction can obtain static attribute data and dynamic behavior data through various public, legal and compliant means, such as obtaining them from public datasets or obtaining them from users with user authorization.

[0023] Step 102: Extract features from static attribute data and dynamic behavior data to obtain feature data of the target user.

[0024] In this embodiment, the target user is the user who participates in the activation event.

[0025] In this embodiment, step 102 includes: first, cleaning, filling in missing information, and encoding the static attribute data to form a user basic attribute vector; then, aggregating and serializing the dynamic behavior data according to a preset time window to extract statistical features such as behavior frequency, duration, interval, conversion rate, and most recent behavior, and further obtaining behavior representation vectors through sequence models / attention mechanisms; finally, concatenating or weighting the two types of feature vectors, and obtaining target user feature data for subsequent model input through dimensionality reduction or feature selection.

[0026] Optionally, static attribute data includes device attribute characteristics, geographic characteristics, and advertising-related characteristics. The advertising-related characteristics include advertising plans, advertising creatives, and channel sources. Step 102 includes: based on historical advertising data, calculating the historical average revenue and / or the historical proportion of high-value users corresponding to the advertising-related characteristics to construct historical quality aggregated characteristics; and using these historical quality aggregated characteristics as part of the target user's feature data. These historical quality aggregated characteristics are prior features derived from historical data statistics. For example, the proportion of high-value users generated by a certain advertising plan or creative in the past 3 days is used to provide initial prediction basis during the cold start phase (when there is no user behavior).

[0027] Step 103: Input the feature data into the pre-trained first prediction model and the second prediction model respectively to obtain the profit prediction value output by the first prediction model and the behavior prediction value output by the second prediction model.

[0028] In this embodiment, the revenue prediction value represents the cumulative revenue of the target user within a preset future period, and the behavior prediction value represents the number of times a specific interactive behavior of the target user occurs within the preset future period. The preset future period can be a period calculated from the activation event, such as 24 hours after activation. Specific interactive behaviors include incentivized video ad exposure.

[0029] In this embodiment, both the first prediction model and the second prediction model can be regression models based on the Extreme Gradient Boosting (XGBoost) algorithm. Optionally, the first prediction model and the second prediction model can also employ tree models such as LightGBM or CatBoost, or multi-task learning models based on deep learning such as DNN or MMoE.

[0030] In this embodiment, the feature data (including user basic attribute features, historical interaction behavior sequence features, content / scene context features, and statistical aggregation features, etc.) constructed for the target user at the current moment is preprocessed and quantized, and then fed as a unified input into the first prediction model and the second prediction model that have been trained offline in parallel for forward inference. The first prediction model outputs a revenue prediction value, which is used to represent the cumulative revenue of the target user in the future preset period, and the second prediction model outputs a behavior prediction value, which is used to represent the number of times the target user's specific interaction behavior occurs in the future preset period. The two outputs can be further used for subsequent user segmentation, strategy selection, or resource allocation decisions.

[0031] Step 104: Match the revenue forecast, behavior forecast, and pre-configured joint feedback conditions to obtain the matching results.

[0032] In this embodiment, the joint postback condition refers to a condition that simultaneously includes dual constraints on the user's future lifetime value and specific interactive behaviors (such as the number of incentivized video exposures), and multiple sets of conditions are logically related by OR; satisfying any one of the sub-conditions triggers a postback. Taking an application's delivery configuration as an example, the joint postback condition configuration is shown in Table 1: Table 1

[0033] At any prediction time point (e.g., 35 minutes after activation), when the user's dual-model prediction value first meets any of the preset conditions in Table 1 above, the user is immediately identified as a high-value heavy user, and the user's activation event is sent back to the advertising platform as a positive sample for deep conversion in real time.

[0034] Step 104 above includes: obtaining the current delivery status data of the current advertising delivery task, the current delivery status data including at least one of the delivery stage, current return on investment, current feedback rate, and current consumption scale; dynamically determining the target joint feedback conditions based on the current delivery status data, the target joint feedback conditions including the dynamically adjusted revenue threshold and the dynamically adjusted behavior threshold; determining whether the revenue prediction value and the behavior prediction value are greater than or equal to the dynamically adjusted revenue threshold and the dynamically adjusted behavior threshold in the target joint feedback conditions; if so, it is determined that the joint feedback conditions are met.

[0035] Step 105: In response to the matching result indicating that the joint feedback condition is met, the activation event is fed back to the advertising platform as a target conversion sample for model optimization on the advertising platform.

[0036] In this embodiment, when the matching result indicates that the current activation event meets the preset joint feedback conditions, the activation event, together with the necessary conversion features and identification information, is encapsulated into a target conversion sample and sent back to the advertising platform through a secure interface. After receiving and verifying the sample, the advertising platform writes it into the training / incremental learning dataset to update model parameters such as conversion prediction, bidding, or audience targeting, thereby achieving continuous optimization of the model's performance on the advertising platform.

[0037] In a specific example, the effectiveness of this disclosure was verified through A / B testing. Forty new advertising accounts were created. Twenty accounts were configured with model feedback (experimental group), using the aforementioned model to predict joint decision logic; the other 20 accounts were configured with rule feedback (control group), using rules based on users' actual cumulative behavior for decision-making (the threshold conditions were exactly the same, but the judgment criteria were the user's current actual revenue and actual incentive video exposures). Specific experimental data are shown in Table 2.

[0038] Table 2

[0039] The experiments shown in Table 2 demonstrate that, compared to the rule-based backhaul group, the model-based backhaul group has a 130% higher consumption rate, a significantly lower backhaul latency, and a roughly equal return on investment (only a 1.74% decrease, which is within an acceptable range).

[0040] Compared to the traditional approach of waiting for actual behavior to occur before transmitting data back, this disclosure significantly reduces invalid long-term network polling and long-term sensor monitoring of terminal devices through short-term window feature extraction and joint prediction, thereby reducing the power consumption of terminal devices and reducing the network bandwidth usage for sending delayed data to the server.

[0041] The data feedback method based on dual-model joint prediction provided in the embodiments of this disclosure firstly acquires static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to the activation event of a target application; secondly, it extracts features from the static attribute data and dynamic behavior data to obtain feature data of the target user; thirdly, it inputs the feature data into a pre-trained first prediction model and a second prediction model respectively to obtain the revenue prediction value output by the first prediction model and the behavior prediction value output by the second prediction model, wherein the revenue prediction value represents the cumulative revenue of the target user in a future preset period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs in a future preset period; then, it matches the revenue prediction value and the behavior prediction value with pre-configured joint feedback conditions to obtain a matching result; finally, in response to the matching result indicating that the joint feedback conditions are met, the activation event is fed back as a target conversion sample to the advertising platform for model optimization of the advertising platform. Therefore, through model prediction, high-value activation events can be quickly identified and transmitted back based on user behavior and profiles after user activation, greatly reducing transmission delay; more accurate and timely high-value information can be transmitted back to the media, which can help the advertising platform's model learn and expand the delivery volume more quickly; in the cold start stage where users have no historical behavior or sparse behavior, the model can make effective predictions based on limited features, overcoming the shortcomings of traditional methods that cannot effectively transmit back information in the cold start stage.

[0042] In some optional implementations of this disclosure, the above-mentioned extraction of features from static attribute data and dynamic behavior data to obtain the feature data of the target user includes: constructing time series incremental features and / or rate features within a continuous time window based on dynamic behavior data; and fusing the time series incremental features and / or rate features with the cumulative value features of static attribute data and dynamic behavior data to obtain the feature data of the target user.

[0043] In this optional implementation, the time series incremental feature is a cumulative value feature reflecting the trend of user behavior changes (such as cumulative revenue, cumulative ad exposures, cumulative launches, etc.), and the time series rate feature is a rate-type feature used to reflect the trend of user behavior changes, the speed of behavior growth, and the level of activity.

[0044] Time series rate features include: behavior increments in the last 5 minutes, behavior increments in the last 10 minutes, behavior frequency per unit time, ad exposure growth rate, revenue growth rate, and cross-combination relationships between multiple behaviors. Compared to simply using cumulative value features, time series rate features can identify user value change trends earlier, thereby improving the model's early prediction ability for high-value users.

[0045] In this optional implementation, time series features are constructed within a preset continuous time window based on the dynamic behavior data of the target user. The time series features include incremental features of behavior in adjacent windows and / or rate features of change per unit time. Subsequently, the incremental features and / or rate features are fused with static attribute data and cumulative value features of dynamic behavior data at the feature level (e.g., splicing, weighted combination, or fusion after dimensionality reduction / embedding mapping) to obtain target user feature data that can simultaneously characterize the user's long-term stable attributes, historical cumulative preferences, and short-term behavioral change trends.

[0046] Optionally, the above-mentioned extraction of features from static attribute data and dynamic behavior data to obtain the feature data of the target user further includes: cross-combining the feature of the frequency of specific behaviors in the dynamic behavior data with the feature of application usage duration to calculate the behavior frequency feature per unit duration; and / or, multiplying the feature of the frequency of specific behaviors in the dynamic behavior data with the feature of application usage duration to construct the behavior and usage duration cross feature; and using the above features as part of the feature data of the target user.

[0047] In some optional implementations of this disclosure, the aforementioned static attribute data includes a list of installed applications on the user's terminal device. Extracting features from the static attribute data and dynamic behavior data to obtain the target user's feature data includes: standardizing and de-identifying the application package names in the installed application list, and mapping them using a hash algorithm to obtain a processed application list; converting the processed application list into application vector features using a preset word embedding algorithm, and using the application vector features as part of the target user's feature data.

[0048] In this optional implementation, the application package name standardization and desensitization process involves standardizing the format of the list of applications installed on the user's device (e.g., converting it to com.xxx.video format) and removing sensitive identifiers that may involve user privacy to comply with data compliance requirements.

[0049] In this optional implementation, hashing algorithms and word embedding algorithms are techniques from the field of natural language processing applied to feature engineering. First, the application package name is mapped to a fixed-length numeric symbol using a hashing algorithm. Then, the discrete application list is converted into a low-dimensional dense vector (embedding vector) containing contextual semantic relationships (i.e., the associations between applications) using a word embedding algorithm, which can then be read by the machine learning model.

[0050] In this optional implementation, after obtaining the list of installed applications on the user's terminal device as static attribute data, the package names of each application in the list are first standardized and anonymized to eliminate naming differences and avoid directly exposing the original identification information. Then, the processed package names are mapped using a hash algorithm to generate an irreversible set of application identifiers, thus obtaining the processed application list. On this basis, a preset word embedding algorithm is used to vectorize the processed application list, mapping each application identifier into a low-dimensional dense application vector feature. This application vector feature is used as part of the target user feature data for subsequent user profile construction, behavior modeling, or predictive analysis.

[0051] In some optional implementations of this disclosure, the above-mentioned joint feedback conditions include multiple sets of sub-conditions, each set of sub-conditions including a corresponding revenue threshold and behavior threshold: matching the revenue prediction value and behavior prediction value with the pre-configured joint feedback conditions to obtain the matching result includes: for each set of sub-conditions, determining whether the revenue prediction value is greater than or equal to the revenue threshold in that sub-condition, and whether the behavior prediction value is greater than or equal to the behavior threshold in that set of sub-conditions; If there exists at least one set of sub-conditions that satisfy the predicted revenue value being greater than or equal to the corresponding revenue threshold and the predicted behavior value being greater than or equal to the corresponding behavior threshold, then the joint backhaul condition is determined to be satisfied.

[0052] In this optional implementation, the joint backhaul conditions are pre-configured into a set of multiple sub-conditions, each of which contains a revenue threshold and a behavior threshold. After obtaining the revenue prediction value and behavior prediction value corresponding to the target object, the system sequentially matches and judges each set of sub-conditions, that is, it checks whether the revenue prediction value is greater than or equal to the revenue threshold of the set and whether the behavior prediction value is greater than or equal to the behavior threshold of the set. When at least one set of sub-conditions simultaneously satisfies the above two threshold constraints, the matching result is determined to satisfy the joint backhaul conditions, and the subsequent joint backhaul processing is triggered accordingly.

[0053] Optionally, the revenue threshold and behavior threshold in the multiple sets of sub-conditions are dynamically determined through the following steps: obtaining historical users' real revenue distribution data, real behavior distribution data, and return on investment data after feedback; based on the preset constraints of target feedback scale and target return on investment data, performing quantile statistics on the real revenue distribution data and real behavior distribution data, and determining multiple sets of revenue threshold and behavior threshold combinations that meet the constraints.

[0054] In some optional implementations of this disclosure, the method further includes: responding to the matching result indicating that the joint feedback condition is not met, waiting for a preset time interval, and then obtaining the latest dynamic behavior data of the target user within the updated time window; updating the feature data based on the latest dynamic behavior data to obtain the updated feature data; and inputting the updated feature data into the first prediction model and the second prediction model for a new round of prediction and matching until the joint feedback condition is met or the preset maximum number of predictions is reached.

[0055] In this optional implementation, the total duration corresponding to the maximum number of predictions shall not exceed the future preset period.

[0056] In this optional implementation, when the matching result indicates that the joint feedback condition is not currently met, the system enters an iterative prediction process: First, it waits for a preset time interval, then acquires the latest dynamic behavior data of the target user within the updated time window, and incrementally updates the original feature data based on the latest dynamic behavior data to obtain the updated feature data; next, the updated feature data is input into the first prediction model and the second prediction model respectively to perform a new round of prediction, and the outputs of the two models are matched to determine whether the joint feedback condition is met; if it is still not met, the loop of "waiting—acquiring—updating—predicting—matching" is repeated until the joint feedback condition is met or the number of iterations reaches the preset maximum number of predictions, thereby ensuring real-time performance while avoiding infinite loops and resource consumption.

[0057] In some optional implementations of this disclosure, the first and second prediction models are trained through the following steps: acquiring sample feature data of historically activated users and their corresponding real labels; performing logarithmic transformation on the real labels and the return threshold based on a preset return threshold corresponding to the real labels, and calculating the distance between the transformed real labels and the transformed return threshold; dynamically assigning weights to each training sample according to the distance, wherein a first weight is assigned to training samples with a distance less than a preset distance threshold, and a second weight is assigned to training samples with a distance greater than or equal to the preset distance threshold, wherein the first weight is greater than the second weight; and iteratively training the corresponding initial model using a weighted loss function based on the weights to obtain the trained first or second prediction model, wherein when training the first prediction model, the real label is the real revenue label, and the return threshold is the revenue return threshold; and when training the second prediction model, the real label is the real behavior label, and the return threshold is the behavior return threshold.

[0058] In this optional implementation, the backhaul threshold is the threshold in the joint decision condition, that is, the backhaul threshold corresponds to the revenue threshold or behavior threshold in the joint backhaul condition.

[0059] In this optional implementation, the real label is a supervision signal in the training of the machine learning model. In this disclosure, the real label specifically refers to the total real benefit or the number of real behaviors obtained by strictly waiting for the user to activate for a full period of time (such as 24 hours).

[0060] In this optional implementation, introducing a logarithmic transformation when calculating the distance can reduce the absolute error caused by the long-tailed maxima, making the model pay more attention to the relative error and improving the stability of regression prediction.

[0061] In this optional implementation, the weighted loss function is used to measure the difference between the model's predicted values ​​and the true values. The weighted loss function disclosed here, based on the traditional Huber loss function, dynamically assigns weights based on the distance of the sample's true value from the backpropagation threshold (the closer the distance, the higher the weight), such as the weighted Huber loss function. This allows the model to no longer simply pursue the minimum global error, but rather focus on improving the classification accuracy near the "critical backpropagation boundary."

[0062] In this optional implementation, the weighted Huber loss function increases sample weights when the user's true label is close to the key behavior threshold; it appropriately reduces the weights of extremely high-value samples to minimize the impact of outliers on overall training stability. Compared to the traditional mean squared error loss function, the weighted Huber Loss in this scheme simultaneously considers: regression prediction stability, outlier robustness, and key behavior boundary identification ability, thereby more accurately identifying high-value addictive users and improving the stability of model backpropagation accuracy and return on investment.

[0063] In this optional implementation, a weighted loss function is used to iteratively train the corresponding initial model, including: calculating the weighted Huber loss value based on the dynamically assigned weights and the distance between the calculated transformed true label and the transformed backpropagation threshold; and updating the parameters of the corresponding initial model with the goal of minimizing the weighted Huber loss value.

[0064] In this optional implementation, samples that are close to the backhaul threshold are given higher weights than samples that are not close to the backhaul threshold, thereby improving the model's ability to distinguish near the key backhaul boundary. This is because in the advertising backhaul scenario, the model's ultimate goal is not just to reduce the overall regression error, but to focus on whether user data should be backhauled to the media. Therefore, whether the predicted value crosses the backhaul threshold is more important than whether the overall training error is minimized.

[0065] In this optional implementation, the executing entity first sets the key behavior feedback threshold, for example: the feedback threshold is 3.9 yuan, and performs Log transformation on the real label LTV24H through formula (1).

[0066] y_true_log = log(1 + LTV24H)(1) The distance between the sample and the return threshold is calculated using equation (2): distance = | y_true_log - log(1 + threshold) | (2) In equation (2), distance represents the difference between the true value of the sample and the back-transmission threshold, and threshold represents the back-transmission threshold.

[0067] The sample weights are dynamically adjusted based on distance: Samples near the threshold (e.g., distance < 0.3): weight = 1.5, used to enhance the model's ability to learn key boundary samples; Ordinary samples: weight = 1.0, maintaining the default weights; Extremely high-value long-tail samples (e.g., LTV > threshold × 3): weight = 0.5, appropriately reducing the weights to avoid extreme high-value samples affecting the overall boundary stability.

[0068] In this optional implementation, the first and second prediction models can be trained as follows: First, acquire the sample feature data of historically activated users and their corresponding real labels; then, based on the preset backhaul threshold corresponding to the real labels, perform logarithmic transformation on the real labels and the backhaul threshold respectively, and calculate the distance between them in the logarithmic space; then, dynamically weight each training sample according to the distance, where samples with a distance less than the preset distance threshold are given higher weights, and samples with a distance greater than or equal to the preset distance threshold are given lower weights; finally, introduce the weights into the weighted loss function, and iteratively train the initial model to obtain the trained first or second prediction model. When training the first prediction model, the real label is the real revenue label and the backhaul threshold is the revenue backhaul threshold; when training the second prediction model, the real label is the real behavior label and the backhaul threshold is the behavior backhaul threshold. Through the above dynamic weighting mechanism, the model can focus on optimizing whether the key behavior backhaul standard is met, thereby improving the backhaul precision, backhaul recall, and return on investment stability, and reducing the false backhaul rate.

[0069] Optionally, in this disclosure, the first and second prediction models can be retrained offline on a scheduled basis. The update steps for both are as follows: daily scheduled model training tasks are executed, the latest historical activated user data is automatically retrieved, and the following processes are re-completed: training sample construction, feature generation, parameter search, model evaluation, and comparison of the old and new models. After the model training is completed, the system will automatically evaluate and compare the performance of the new model with the current online model, including metrics such as mean squared error, mean absolute error, backpropagation accuracy, backpropagation recall, and backpropagation F1 score. The system will only automatically replace the online model when the performance of the new model meets preset conditions to avoid model performance degradation.

[0070] In some optional implementations of this disclosure, the above-mentioned dynamic allocation of weights to each training sample based on distance includes: if the true label of a training sample is greater than a preset multiple of the backhaul threshold, then the weight of the training sample is reduced.

[0071] In this optional implementation, training samples whose true labels are greater than a preset multiple of the backhaul threshold are samples whose distance is within a preset distance threshold but whose labels are extremely high.

[0072] In this optional implementation, the initial weights for each training sample are first calculated based on its distance from the backpropagation threshold, and then the magnitude constraint of the true label is introduced for dynamic correction: when the true label value of a training sample is greater than a preset multiple of the backpropagation threshold, it is determined that the sample may belong to an extreme value or an abnormal distribution, and its weight is reduced (e.g., reduced by a preset reduction coefficient or decreased proportionally) to reduce the contribution of the sample to the loss function and gradient backpropagation; the remaining samples maintain their original weights or are normally assigned weights according to the distance rule, thereby ensuring the effective learning effect of effective samples while suppressing the adverse effects of large-label samples on the stability of model training.

[0073] In some optional implementations of this disclosure, the sample feature data of the historically activated users is determined through the following steps: obtaining the original data of historically activated users within a preset total historical time period; dividing the preset total historical time period into multiple candidate training time windows with different time spans; training multiple candidate models based on the data within each candidate training time window; evaluating the error indices of the multiple candidate models on the same validation set, and selecting the candidate training time window corresponding to the candidate model with the smallest error index as the target training time window; obtaining the original data of historically activated users within the target training time window, and obtaining the sample feature data of historically activated users based on the original data.

[0074] In this optional implementation, the candidate training time window refers to the historical time span of the training data (e.g., 7 days or 14 days). The candidate training time window breaks the limitation of traditional models that use the "most recent N days" of data for training (N>1). By dividing historical data into different time spans (e.g., the most recent 7 days, 14 days, 30 days), candidate models are trained separately, and the most suitable data time range is automatically searched through the validation set error to improve the model's generalization ability.

[0075] In this optional implementation, multiple candidate models are trained separately, and then uniformly evaluated on the same validation set. The model with the lowest mean squared error is selected as the optimal model for this training.

[0076] In this optional implementation, the raw data of historically active users within a preset total historical time period are first obtained, and this total historical time period is divided into multiple candidate training time windows according to different time spans. Then, data within each candidate training time window is extracted to train multiple candidate models, and the error index of each candidate model is evaluated on the same validation set. Based on the evaluation results, the candidate training time window corresponding to the candidate model with the smallest error index is selected as the target training time window. Finally, the historically active user data within the target training time window is obtained and used as sample feature data of historically active users for subsequent modeling or prediction.

[0077] like Figure 2 As shown, after a user clicks on an advertisement and activates the target application, the execution entity collects static attribute and dynamic behavior data in parallel by gathering user data. After processing by the feature processing layer, the data is input into the first and second prediction models for parallel prediction. Prediction begins 15 minutes after activation, with each cycle lasting 5 minutes. When the predicted value of the two models meets the preset joint feedback conditions (i.e., matches the joint feedback conditions), feedback is triggered. During this feedback process, high-value information (high-value users) is identified and fed back to the advertising platform, where the model is optimized. If the conditions are not met, the system waits 5 minutes before starting the next prediction cycle.

[0078] Further reference Figure 3 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a data backhaul device based on dual-model joint prediction. This device embodiment is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0079] like Figure 3As shown, the data feedback device 300 based on dual-model joint prediction provided in this embodiment includes: an acquisition unit 301, an extraction unit 302, an input unit 303, a matching unit 304, and a feedback unit 305. The acquisition unit 301 can be configured to acquire static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to a target application activation event. The extraction unit 302 can be configured to extract features from the static attribute data and dynamic behavior data to obtain feature data of the target user. The input unit 303 can be configured to input the feature data into a pre-trained first prediction model and a second prediction model respectively to obtain a revenue prediction value output by the first prediction model and a behavior prediction value output by the second prediction model. The revenue prediction value represents the cumulative revenue of the target user within a preset future period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs within the preset future period. The matching unit 304 can be configured to match the revenue prediction value and the behavior prediction value with pre-configured joint feedback conditions to obtain a matching result. The aforementioned feedback unit 305 can be configured to, in response to a matching result indicating that the joint feedback conditions are met, send the activation event as a target conversion sample back to the advertising platform for model optimization on the advertising platform.

[0080] In this embodiment, the specific processing of the data feedback device 300 based on dual-model joint prediction, including the acquisition unit 301, extraction unit 302, input unit 303, matching unit 304, and feedback unit 305, and the resulting technical effects, can be found in the following references: Figure 1 The relevant descriptions of steps 101, 102, 103, 104, and 105 in the corresponding embodiments will not be repeated here.

[0081] In some optional implementations of this disclosure, the extraction unit 302 is configured to: construct time series incremental features and / or rate features within a continuous time window based on dynamic behavioral data; and fuse the time series incremental features and / or rate features, as well as the cumulative value features of static attribute data and dynamic behavioral data, to obtain the feature data of the target user.

[0082] In some optional implementations of this disclosure, the aforementioned static attribute data includes a list of installed applications on the user terminal device. The aforementioned extraction unit 302 is configured to: standardize and de-identify the application package names in the list of installed applications, and map them using a hash algorithm to obtain a processed application list; convert the processed application list into application vector features using a preset word embedding algorithm, and use the application vector features as part of the target user's feature data.

[0083] In some optional implementations of this disclosure, the aforementioned joint feedback conditions include multiple sets of sub-conditions, each set of sub-conditions including a corresponding revenue threshold and a behavior threshold. The matching unit 304 is configured to: for each set of sub-conditions, determine whether the predicted revenue value is greater than or equal to the revenue threshold in that sub-condition, and whether the predicted behavior value is greater than or equal to the behavior threshold in that set of sub-conditions; if at least one set of sub-conditions satisfies that the predicted revenue value is greater than or equal to the corresponding revenue threshold and the predicted behavior value is greater than or equal to the corresponding behavior threshold, then it is determined that the joint feedback conditions are met.

[0084] In some embodiments of this disclosure, the data feedback device 300 based on dual-model joint prediction further includes: a prediction unit (not shown in the figure), which is configured to: in response to a matching result indicating that the joint feedback condition is not met, wait for a preset time interval, obtain the latest dynamic behavior data of the target user within the updated time window; update the feature data based on the latest dynamic behavior data to obtain updated feature data; input the updated feature data into the first prediction model and the second prediction model for a new round of prediction and matching until the joint feedback condition is met or the preset maximum number of predictions is reached.

[0085] In some optional implementations of this disclosure, the first prediction model and the second prediction model are trained by a training unit (not shown in the figure), which is configured to: acquire sample feature data of historically activated users and their corresponding real labels; perform logarithmic transformation on the real labels and the return threshold based on a preset return threshold corresponding to the real labels, and calculate the distance between the transformed real labels and the transformed return threshold; dynamically assign weights to each training sample according to the distance, wherein a first weight is assigned to training samples whose distance is less than a preset distance threshold, and a second weight is assigned to training samples whose distance is greater than or equal to the preset distance threshold, wherein the first weight is greater than the second weight; and iteratively train the corresponding initial model using a weighted loss function based on the weights to obtain the trained first prediction model or the second prediction model, wherein when training the first prediction model, the real label is the real revenue label, and the return threshold is the revenue return threshold; when training the second prediction model, the real label is the real behavior label, and the return threshold is the behavior return threshold.

[0086] In some optional implementations of this disclosure, the training device is further configured to reduce the weight of the training sample if the true label of the training sample is greater than a preset multiple of the backhaul threshold.

[0087] In some optional implementations of this disclosure, the training device is further configured to: acquire the original data of historically activated users within a preset total historical time period; divide the preset total historical time period into multiple candidate training time windows with different time spans; train multiple candidate models based on the data within each candidate training time window; evaluate the error indices of the multiple candidate models on the same validation set, and select the candidate training time window corresponding to the candidate model with the smallest error index as the target training time window; acquire the original data of historically activated users within the target training time window, and obtain sample feature data of historically activated users based on the original data.

[0088] The data feedback device based on dual-model joint prediction provided in the embodiments of this disclosure firstly acquires static attribute data and dynamic behavior data of a user terminal device within a preset initial time window due to a target application activation event. Secondly, the extraction unit 302 extracts features from the static attribute data and dynamic behavior data to obtain feature data of the target user. Thirdly, the input unit 303 inputs the feature data into a pre-trained first prediction model and a second prediction model respectively to obtain the revenue prediction value output by the first prediction model and the behavior prediction value output by the second prediction model. The revenue prediction value represents the cumulative revenue of the target user within a preset future period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs within the preset future period. Then, the matching unit 304 matches the revenue prediction value and the behavior prediction value with pre-configured joint feedback conditions to obtain a matching result. Finally, in response to the matching result indicating that the joint feedback conditions are met, the feedback unit 305 sends the activation event as a target conversion sample back to the advertising platform for model optimization. By using dual-model advance prediction, the feedback delay of high-value users is reduced, and the learning efficiency of the advertising model is improved.

[0089] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0090] Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their patterns are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0091] like Figure 4 As shown, the electronic device 400 includes a computing unit 401, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0092] Multiple components in electronic device 400 are connected to I / O interface 405, including: interaction unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of displays, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows electronic device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0093] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as the data backhaul method based on dual-model joint prediction. For example, in some embodiments, the data backhaul method based on dual-model joint prediction can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the data backhaul method based on dual-model joint prediction described above can be performed. Alternatively, in other embodiments, computing unit 401 may be configured by any other suitable means (e.g., by means of firmware) to perform a data backhaul method based on dual-model joint prediction.

[0094] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0095] The program code used to implement the methods of this disclosure can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data return device based on dual-model joint estimation, such that when executed by the processor or controller, the program code enables the patterns / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0096] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0097] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0098] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0099] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0100] The foregoing description of specific exemplary embodiments of this disclosure is for illustrative and explanatory purposes. These descriptions are not intended to limit this disclosure to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of this disclosure and their practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of this disclosure, as well as various different choices and variations. The scope of this disclosure is intended to be defined by the claims and their equivalents.

[0101] The above are merely embodiments of this disclosure and are not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this disclosure should be included within the protection scope of this disclosure.

Claims

1. A data backhaul method based on dual-model joint prediction, characterized in that, The method includes: Acquire static attribute data and dynamic behavior data of the user terminal device within a preset initial time window due to the activation event of the target application; Extract the features from the static attribute data and the dynamic behavior data to obtain the feature data of the target user; The feature data is input into a pre-trained first prediction model and a second prediction model respectively to obtain the revenue prediction value output by the first prediction model and the behavior prediction value output by the second prediction model. The revenue prediction value represents the cumulative revenue of the target user in a future preset period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs in the future preset period. The predicted revenue and the predicted behavior are matched with the pre-configured joint feedback conditions to obtain the matching result; In response to the matching result indicating that the joint feedback condition is met, the activation event is fed back to the advertising platform as a target conversion sample for model optimization on the advertising platform.

2. The method according to claim 1, characterized in that, The step of extracting features from the static attribute data and the dynamic behavior data to obtain the target user's feature data includes: Based on the dynamic behavior data, time series incremental features and / or rate features are constructed within continuous time windows; The feature data of the target user is obtained by fusing the time series incremental features and / or rate features, as well as the cumulative value features of the static attribute data and the dynamic behavior data.

3. The method according to claim 2, characterized in that, The static attribute data includes a list of installed applications on the user terminal device. Extracting features from the static attribute data and the dynamic behavior data to obtain the target user's feature data includes: The application package names in the installed application list are standardized and de-identified, and then mapped using a hash algorithm to obtain the processed application list; The processed application list is converted into application vector features using a preset word embedding algorithm, and these application vector features are used as part of the target user's feature data.

4. The method according to claim 1, characterized in that, The joint backhaul conditions include multiple sets of sub-conditions, each set of sub-conditions including a corresponding revenue threshold and a behavior threshold. The matching of the predicted revenue value and the predicted behavior value with the pre-configured joint backhaul conditions to obtain the matching result includes: For each of the multiple sets of sub-conditions, determine whether the predicted revenue value is greater than or equal to the revenue threshold in that sub-condition, and whether the predicted behavior value is greater than or equal to the behavior threshold in that set of sub-conditions. If there exists at least one set of sub-conditions that satisfy the predicted revenue value being greater than or equal to the corresponding revenue threshold and the predicted behavior value being greater than or equal to the corresponding behavior threshold, then the joint backhaul condition is determined to be satisfied.

5. The method according to claim 1, characterized in that, The method further includes: In response to the matching result indicating that the joint feedback condition is not met, after waiting for a preset time interval, the latest dynamic behavior data of the target user within the updated time window is obtained. Based on the latest dynamic behavior data, the feature data is updated to obtain the updated feature data; The updated feature data is input into the first prediction model and the second prediction model for a new round of prediction and matching until the joint backhaul condition is met or the preset maximum number of predictions is reached.

6. The method according to any one of claims 1-5, characterized in that, The first prediction model and the second prediction model are obtained through the following steps: Obtain sample feature data of historically activated users and their corresponding real labels; Based on a preset return threshold corresponding to the real tag, logarithmic transformation is performed on the real tag and the return threshold respectively, and the distance between the transformed real tag and the transformed return threshold is calculated. Weights are dynamically assigned to each training sample based on the distance, wherein a first weight is assigned to training samples whose distance is less than a preset distance threshold, and a second weight is assigned to training samples whose distance is greater than or equal to the preset distance threshold, wherein the first weight is greater than the second weight. Based on the weights, a weighted loss function is used to iteratively train the corresponding initial model to obtain the first prediction model or the second prediction model after training. When training the first prediction model, the true label is the true revenue label and the feedback threshold is the revenue feedback threshold. When training the second prediction model, the true label is the true behavior label and the feedback threshold is the behavior feedback threshold.

7. The method according to claim 6, characterized in that, The step of dynamically assigning weights to each training sample based on the distance includes: If the true label of the training sample is greater than a preset multiple of the backhaul threshold, then the weight of the training sample is reduced.

8. The method according to claim 6, characterized in that, The sample feature data of the historically activated users are determined through the following steps: Obtain the raw data of historically activated users within a preset total historical time period; The preset total historical time period is divided into multiple candidate training time windows with different time spans; Multiple candidate models are obtained by training based on the data within each of the candidate training time windows; The error metrics of the multiple candidate models are evaluated on the same validation set, and the candidate training time window corresponding to the candidate model with the smallest error metric is selected as the target training time window. Obtain the original data of historically activated users within the target training time window, and obtain the sample feature data of the historically activated users based on the original data.

9. A data backhaul device based on dual-model joint prediction, characterized in that, The device includes: The acquisition unit is configured to acquire static attribute data and dynamic behavior data of the user terminal device within a preset initial time window due to the activation event of the target application; The extraction unit is configured to extract features from the static attribute data and the dynamic behavior data to obtain feature data of the target user. The input unit is configured to input the feature data into a pre-trained first prediction model and a second prediction model respectively to obtain a revenue prediction value output by the first prediction model and a behavior prediction value output by the second prediction model, wherein the revenue prediction value represents the cumulative revenue of the target user in a future preset period, and the behavior prediction value represents the number of times a specific interaction behavior of the target user occurs in the future preset period. The matching unit is configured to match the revenue prediction value, the behavior prediction value, and pre-configured joint feedback conditions to obtain a matching result; The feedback unit is configured to, in response to the matching result indicating that the joint feedback condition is met, send the activation event as a target conversion sample back to the advertising platform for model optimization of the advertising platform.

10. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

11. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-8.

Images