An advertising delivery method, apparatus, equipment, and storage medium

By performing dimensionless processing on advertising delivery characteristics, using attention networks and gating value calculations, and combining gradient boosting decision processing, an estimated return on investment (ROI) is generated, solving the problem of inaccurate ROI estimation in existing technologies and achieving more accurate advertising delivery operations.

CN123071035APending Publication Date: 2026-11-10GUANGZHOU LEWAN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611297460.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing advertising platforms, the accuracy of return on investment (ROI) estimation is low. The manual statistics and verification methods lead to inaccurate ROI estimation, and the lack of quantitative assessment of the degree of abnormality of feature values ​​results in distorted estimation results and operational delays.

Method used

By obtaining multidimensional delivery features of target ads and performing dimensionless processing, enhanced feature vectors are calculated using attention networks and gating values. Combined with gradient boosting decision processing, an estimated return on investment (ROI) is generated, and delivery control operations are executed based on the estimated ROI.

Benefits of technology

It significantly improves the accuracy of return on investment estimation and advertising placement operations, reduces the interference of abnormal feature values ​​on estimation, and achieves adaptive assessment of feature importance and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071035A_ABST
    Figure CN123071035A_ABST
Patent Text Reader

Abstract

This application discloses an advertising delivery method, apparatus, device, and storage medium, comprising: inputting the first multi-dimensional features of a target advertisement into a fully connected layer for nonlinear transformation to obtain a latent representation vector; performing attention calculation on the latent representation vector to obtain attention weights for each feature dimension; determining a corresponding gate value based on the deviation of the current value of each feature dimension from its value distribution; multiplying the value of each dimension, attention weight, and gate value element-wise in the latent representation vector to obtain an enhanced feature vector; concatenating the multi-dimensional delivery features and the enhanced feature vector along the feature dimensions to obtain a fused feature set, and performing a first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment (ROI) of the target advertisement; and executing delivery control operations based on the estimated ROI. This solution addresses the problem of low estimation accuracy by obtaining the estimated ROI based on a fused feature set that integrates the enhanced feature vector and the multi-dimensional delivery features, thereby improving estimation accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of advertising delivery technology, and in particular to an advertising delivery method, apparatus, device and storage medium. Background Technology

[0002] With the maturity of the mobile internet ecosystem and the rapid growth of the casual and mini-game sectors, advertising has become a core operational method for game developers to acquire users, realize commercial monetization, and maintain product lifecycles, playing an irreplaceable and crucial role throughout the entire game lifecycle.

[0003] In existing technologies, advertising platforms typically provide a built-in data backend that collects advertising metrics such as spend, impressions, and clicks through a statistics module, and presents these metrics to advertisers in the form of reports. Advertisers then manually estimate the ROI (return on investment) based on the spend and conversion data in the reports, and manually execute actions such as continuing to run the campaign, reducing the budget, or shutting it down.

[0004] The current method of manual statistics and verification results in relatively low accuracy in estimating the return on investment. Summary of the Invention This application provides an advertising placement method, apparatus, device, and storage medium that can solve the technical problem of low accuracy in ROI estimation during advertising placement. By performing attention and gating feature enhancement on the first multi-dimensional feature to obtain an enhanced feature vector, and fusing the enhanced feature vector and multi-dimensional placement features to obtain a fused feature set, gradient boosting decision processing is performed based on the fused feature set to obtain the corresponding estimated ROI, which significantly improves the accuracy of ROI estimation and thus improves the accuracy of advertising placement operations.

[0005] In a first aspect, embodiments of this application provide an advertising delivery method, including: The multidimensional delivery features of the target advertisement are obtained, and the multidimensional delivery features are dedimensionalized to obtain the first multidimensional feature. The first multidimensional feature is input into at least one fully connected layer for nonlinear transformation to obtain the latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multidimensional feature. Attention weights for each feature dimension are obtained by performing attention calculations on the latent representation vector using an attention network. Obtain the value distribution of each feature dimension in multiple historical ads of the same type as the target ad, and determine the gate value corresponding to each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The feature dimension whose current value deviates further from its value distribution has a lower gate value. The enhanced feature vector is obtained by multiplying the value of each dimension of the latent representation vector, the attention weight of the corresponding dimension, and the gating value of the corresponding dimension element by element. The multi-dimensional delivery features and the enhanced feature vectors are concatenated along the feature dimensions to obtain a fused feature set. The fused feature set is then subjected to the first gradient boosting decision processing to obtain the estimated return on investment of the target advertisement. Execute placement control operations for target advertisements based on the estimated return on investment.

[0006] Furthermore, the multidimensional projection features are dedimensionalized to obtain the first multidimensional feature, which includes: Identify the data type of each feature in the multidimensional delivery features, keep the original values ​​of numerical features unchanged, and map categorical features to numerical scalars through an embedding layer to obtain a standardized feature set; The standardized feature set is normalized to obtain the first multidimensional feature.

[0007] Furthermore, the fused feature set is used for first-gradient improvement decision processing to obtain the estimated ROI of the target advertisement, including: The first gradient boosting decision processing is performed sequentially on multiple preset time nodes in order from the nearest to the farthest time node to obtain the estimated output ratio corresponding to each preset time node. The processing of the first preset time node takes the fused feature set as input, and the processing of each subsequent preset time node takes the feature set obtained by concatenating the fused feature set with the estimated output ratio of the previously processed preset time node as input, so that the estimated output ratio of the subsequent preset time node is conditional on the estimated results of the previous preset time node. Accordingly, based on the estimated return on investment, the target advertising campaign is subject to delivery control measures, including: The confidence level of the trend direction is calculated based on the consistency of the change direction of the estimated output ratio between adjacent preset time nodes. The confidence level of the trend direction is the proportion of the number of intervals where the change direction between adjacent preset time nodes is the same to the total number of intervals. The deployment control operation is determined based on the trend direction of the time-series trajectory formed by the confidence level of the trend direction and the estimated input-output ratio at each preset time node; When the time series trajectory shows an upward trend and the estimated return on investment at the most recent preset time node is greater than or equal to the first preset threshold, and the confidence level of the trend direction is greater than or equal to the preset direction confidence threshold, continue the deployment operation; when the time series trajectory shows an upward trend but the estimated return on investment at the most recent preset time node is less than the first preset threshold, reduce the budget operation; when the time series trajectory shows an upward trend and the estimated return on investment at the most recent preset time node is greater than or equal to the first preset threshold, but the confidence level of the trend direction is less than the preset direction confidence threshold, reduce the budget operation and extend the data observation period; when the time series trajectory shows a downward trend or a stable trend and the estimated return on investment at the most recent preset time node is less than the second preset threshold, shut down the deployment operation; when the time series trajectory shows a downward trend or a stable trend and the estimated return on investment at the most recent preset time node is greater than or equal to the second preset threshold, reduce the budget operation.

[0008] Furthermore, based on the estimated ROI, the target advertising campaign is subject to delivery control measures, including: The current campaign stage of the target ad is determined based on the percentage of the target ad's cumulative spending budget to the preset total budget. When the spending percentage is lower than the first spending percentage threshold, the current campaign stage is determined to be the first stage. When the spending percentage is between the first spending percentage threshold and the second spending percentage threshold, the current campaign stage is determined to be the second stage. When the spending percentage is higher than the second spending percentage threshold, the current campaign stage is determined to be the third stage. The upper and lower limits of the return on investment ratio are determined according to the current deployment stage. Specifically, the upper and lower limits of the return on investment ratio in the first stage are lower than the corresponding thresholds in the second stage, and the upper and lower limits of the return on investment ratio in the third stage are higher than the corresponding thresholds in the second stage. When the estimated return on investment is greater than or equal to the upper limit threshold of the return on investment corresponding to the current stage of the campaign, the campaign will continue to run for the target ad. When the estimated return on investment (ROI) is less than the upper limit threshold of ROI for the current campaign stage but greater than or equal to the lower limit threshold of ROI for the current campaign stage, the budget reduction operation is performed on the target ad. When the estimated return on investment (ROI) is lower than the lower limit threshold of ROI for the current campaign stage, the target ad will be shut down.

[0009] Furthermore, after obtaining the estimated ROI of the target advertisement, it also includes: When the target ad is in the cold start stage with sparse data, the multi-dimensional delivery features are processed in a second gradient to improve decision-making and obtain an auxiliary return on investment. The estimated return on investment (ROI) and the auxiliary ROI are weighted and combined to obtain the target ROI. The weight of the estimated ROI gradually increases as the cumulative data volume of the target advertisement increases, while the weight of the auxiliary ROI gradually decreases as the cumulative data volume increases. Accordingly, based on the estimated return on investment, the target advertising campaign is subject to delivery control measures, including: Execute delivery control operations for target ads based on the target return on investment.

[0010] Furthermore, based on the estimated ROI, the target advertising campaign is subject to delivery control measures, including: When the estimated return on investment is lower than the preset optimization trigger threshold, the attention weights corresponding to each feature dimension of the attention network output in the same forward processing that generates the estimated return on investment are obtained. Based on the frequency of use and splitting gain of each feature dimension in the first gradient boosting decision processing when splitting tree nodes, the marginal contribution of each feature dimension to the estimated return on investment is determined. The feature dimension with the largest negative marginal contribution is identified as the root cause feature. Based on the covariance of the attention weights of each feature dimension on multiple historical samples of the target advertisement, the associated feature with the highest attention weight covariance with the root cause feature is identified from all feature dimensions except the root cause feature. Based on root cause and correlation characteristics, determine the delivery parameters to be adjusted in the budget reduction operation and the direction of adjustment for each delivery parameter, and then execute the budget reduction operation.

[0011] Furthermore, the gate value for each feature dimension is determined based on the degree of deviation of the current value of each feature dimension from its value distribution, including: Determine the current cumulative data volume of each feature dimension in the multi-dimensional delivery feature. For feature dimensions whose cumulative data volume has not reached the preset data volume threshold, mark them as under-observed dimensions. Based on the degree of deviation of the current value of each feature dimension from its value distribution, calculate the initial gate value corresponding to each feature dimension; For a feature dimension marked as underobserved, its initial gate value is multiplied by a decay coefficient to obtain the target gate value of the feature dimension. The decay coefficient is positively correlated with the ratio of the cumulative data volume of the feature dimension to the preset data volume threshold and is less than one. For feature dimensions that are not labeled as underobserved dimensions, their initial gating values ​​are used as the target gating values ​​for the feature dimensions.

[0012] In a second aspect, embodiments of this application provide an advertising delivery device, comprising: The feature acquisition module is used to acquire multi-dimensional delivery features of the target advertisement; The dimensionless processing module is used to perform dimensionless processing on the multidimensional projection features to obtain the first multidimensional feature; The latent representation vector calculation module is used to input the first multidimensional feature into at least one fully connected layer and perform a nonlinear transformation to obtain the latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multidimensional feature. The attention calculation module is used to perform attention calculation on the latent representation vector through the attention network to obtain the attention weights corresponding to each feature dimension. The gate value determination module is used to obtain the value distribution of each feature dimension in multiple historical ads of the same type as the target ad. Based on the degree of deviation of the current value of each feature dimension from its value distribution, the gate value corresponding to each feature dimension is determined. The feature dimension whose current value deviates further from the value distribution has a lower gate value. The feature enhancement module is used to multiply the value of each dimension of the latent representation vector, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension element by element to obtain the enhanced feature vector. The feature fusion module is used to concatenate multi-dimensional delivery features and enhanced feature vectors along the feature dimensions to obtain a fused feature set, and then perform first-gradient boosting decision processing on the fused feature set to obtain the estimated return on investment of the target advertisement. The campaign execution determination module is used to perform campaign control operations on target ads based on the estimated return on investment.

[0013] In a third aspect, embodiments of this application provide an advertising delivery device, including: Memory and one or more processors; Memory, used to store one or more programs; When one or more programs are executed by one or more processors, the one or more processors implement the advertising delivery method as described in the first aspect.

[0014] In a fourth aspect, embodiments of this application provide a storage medium for storing computer-executable instructions, which, when executed by a computer processor, are used to perform the advertising delivery method as described in the first aspect.

[0015] This application embodiment obtains multi-dimensional delivery features of the target advertisement during ad delivery, performs dimensionless processing on the multi-dimensional delivery features to obtain a first multi-dimensional feature, inputs the first multi-dimensional feature into at least one fully connected layer for nonlinear transformation to obtain a latent representation vector; performs attention calculation on the latent representation vector through an attention network to obtain the attention weight corresponding to each feature dimension; obtains the value distribution of each feature dimension in multiple historical advertisements of the same type as the target advertisement, and determines the gate value corresponding to each feature dimension based on the deviation of the current value of each feature dimension from its value distribution; multiplies the value of each dimension in the latent representation vector, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension element-wise to obtain an enhanced feature vector; concatenates the multi-dimensional delivery features and the enhanced feature vector along the feature dimensions to obtain a fused feature set, and performs a first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment (ROI) of the target advertisement; and performs delivery control operations on the target advertisement based on the estimated ROI. By employing the aforementioned techniques, attention weights are used to learn which features are more important for ROI prediction from the latent representation of feature interactions. Gating values ​​are used to assess which features are abnormal or unreliable in the current sample based on the deviation of the original values ​​from the historical distribution. These two methods, starting from different information sources and following independent computational paths, measure the importance and reliability of features respectively. Then, element-wise multiplication is used to jointly apply this to the latent representation vector to obtain an enhanced feature vector. This strengthens important and reliable features while suppressing unimportant or unreliable features. Simultaneously, the enhanced feature vector is concatenated with the original multidimensional delivery features, enabling the first gradient boosting decision processing to utilize both the enhanced high-quality features and the original features that retain the original distribution information. This avoids the technical problem of low ROI estimation accuracy in existing technologies. In this embodiment, the first gradient boosting decision processing outputs the corresponding predicted ROI, significantly improving the accuracy of ROI estimation. Subsequently, delivery control operations are performed on the target advertisement based on the predicted ROI, further improving the accuracy of the advertisement delivery operation.

[0016] The beneficial effects of the advertising delivery device, advertising delivery equipment, and storage medium provided above can be referenced to the beneficial effects of the advertising delivery method. Attached Figure Description

[0017] Figure 1 This is a flowchart of an advertising delivery method provided in an embodiment of this application; Figure 2 This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 3 This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 4 This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 5This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 6 This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 7 This is a flowchart of another advertising delivery method provided in the embodiments of this application; Figure 8 This is a schematic diagram of the structure of an advertising delivery device provided in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of an advertising delivery device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, specific embodiments of this application will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it. It should also be noted that, for ease of description, only the parts relevant to this application are shown in the drawings, not all of them. Before discussing exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0019] In existing technologies, advertising platforms typically provide built-in data backends that collect ad metrics such as spend, impressions, and clicks through statistical modules, presenting these metrics to advertisers in report form. Advertisers manually estimate ROI based on the spend and conversion data in these reports and then manually execute actions such as continuing ad campaigns, reducing budgets, or shutting down the platform. Current solutions primarily rely on human experience or simple statistical mean extrapolation for ad feature processing, failing to adaptively enhance the original features. For example, when features contain noise or redundant dimensions weakly correlated with ROI, these useless features also participate in ROI estimation, leading to poor adaptability and large fluctuations in accuracy across different ads or campaign stages. Furthermore, ad data is inherently noisy; the same ad's metrics may exhibit abnormal values ​​at different times due to external factors such as audience fluctuations and market competition. Existing solutions lack a quantitative assessment mechanism for the degree of feature value anomalies, allowing outliers to directly enter the ROI estimation process, resulting in distorted estimation results. Moreover, most existing solutions stop at data display or manual estimation, failing to directly drive campaign control operations with the estimation results, resulting in delays and human biases in the data-to-decision loop.

[0020] Based on this, the present application provides an advertising delivery method, apparatus, device, and storage medium. During advertising delivery, the method involves: acquiring multi-dimensional delivery features of the target advertisement; performing dimensionless processing on the multi-dimensional delivery features to obtain a first multi-dimensional feature; inputting the first multi-dimensional feature into at least one fully connected layer for nonlinear transformation to obtain a latent representation vector; performing attention calculation on the latent representation vector through an attention network to obtain the attention weights corresponding to each feature dimension; acquiring the value distribution of each feature dimension in multiple historical advertisements of the same type as the target advertisement; determining the gate value corresponding to each feature dimension based on the deviation of the current value of each feature dimension from its value distribution; multiplying the value of each dimension in the latent representation vector, the corresponding attention weight, and the corresponding gate value element-wise to obtain an enhanced feature vector; concatenating the multi-dimensional delivery features and the enhanced feature vector along the feature dimensions to obtain a fused feature set; performing a first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment (ROI) of the target advertisement; and performing delivery control operations on the target advertisement based on the estimated ROI. By employing the aforementioned techniques, attention weights are used to learn which features are more important for ROI prediction from the latent representation of feature interactions. Gating values ​​are used to assess which features are abnormal or unreliable in the current sample based on the deviation of the original values ​​from the historical distribution. These two methods, starting from different information sources and following independent computational paths, measure the importance and reliability of features respectively. Then, element-wise multiplication is used to jointly apply the enhanced feature vector to the latent representation vector, thus strengthening important and reliable features while suppressing unimportant or unreliable features. Simultaneously, the enhanced feature vector is concatenated with the original multidimensional delivery features, enabling the first gradient boosting decision processing to utilize both the enhanced high-quality features and the original features that retain the original distribution information. This addresses the technical problem in existing technologies where all features are treated equally and noisy features also participate in the estimation, leading to low ROI prediction accuracy. This embodiment significantly improves the accuracy of ROI prediction by outputting the corresponding predicted ROI through the first gradient boosting decision processing. Subsequently, delivery control operations are performed on the target advertisement based on the predicted ROI, thereby improving the accuracy of the advertisement delivery operation.

[0021] Figure 1 This is a flowchart illustrating an advertising delivery method provided in an embodiment of this application. The advertising delivery method provided in this embodiment can be executed by an advertising delivery device, which can be implemented through software and / or hardware. The advertising delivery device can consist of two or more physical entities, or it can consist of a single physical entity. Generally, the advertising delivery device can be an electronic device, such as a PC (personal computer), interactive whiteboard, tablet computer, smart TV, learning machine, or mobile phone.

[0022] The following description uses PC as the primary method for ad delivery as an example. (Refer to...) Figure 1 The specific methods for placing advertisements include: S11. Obtain the multi-dimensional delivery features of the target advertisement, perform dimensionless processing on the multi-dimensional delivery features to obtain the first multi-dimensional feature, input the first multi-dimensional feature into at least one fully connected layer for nonlinear transformation to obtain the latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multi-dimensional feature.

[0023] In one embodiment, when placing an ad, the multi-dimensional placement features of the target ad are first obtained. These features are then dedimensionalized to obtain a first multi-dimensional feature. This first multi-dimensional feature is then input into at least one fully connected layer for nonlinear transformation to obtain a latent representation vector. The multi-dimensional placement features refer to multi-dimensional indicator data collected from multiple data sources to describe the placement status of the target ad and user conversion performance. The target ad refers to the ad campaign for which placement control decisions are currently being made; an ad campaign has an independent budget, bidding strategy, and placement targeting. The "dimensionality" of the multi-dimensional placement features refers to the number of feature columns. For example, if six indicators are collected: spending amount, impressions, clicks, user activations, user retention rate, and user payment amount, then the dimension d=6. A fully connected layer is a network layer in a neural network where all adjacent neurons have connection weights; each input neuron has a learnable weight parameter between it and each output neuron. Dimensionless transformation refers to the normalization transformation of the values ​​of each feature dimension, eliminating the differences in dimensions and magnitudes between different features, so that the values ​​of each feature are mapped to a uniform numerical range. The first multidimensional feature refers to the feature set obtained after dimensionless transformation, where the values ​​of each feature are all within a uniform numerical range, and its dimension is the same as that of the multidimensional projection feature. The latent representation vector refers to the intermediate representation vector obtained after the first multidimensional feature undergoes a nonlinear transformation through a fully connected layer. Its dimension is the same as that of the input first multidimensional feature, but the value of each dimension is no longer a simple linear combination of the original features, but a nonlinear expression that integrates information from multiple original features.

[0024] For example, performance data of the target ad during the current campaign period is obtained from the data interface of the advertising platform, and user behavior data is obtained from the data interface of the application promoted by the target ad. This data is then concatenated to form the original multi-dimensional campaign features. Next, the multi-dimensional campaign features are dedimensionalized to obtain the first multi-dimensional feature. Then, the first multi-dimensional feature is input into at least one pre-constructed fully connected layer. The weight matrix in the fully connected layer linearly combines the input features and then transforms them through a non-linear activation function (such as the ReLU function), outputting a latent representation vector. Each dimension in this latent representation vector no longer corresponds to a specific original feature, but rather is a comprehensive expression of multiple original features after weighted combination and non-linear mapping.

[0025] For example, suppose the original multidimensional delivery features of a target ad are [spending = 5000, impressions = 100000, clicks = 3000, activations = 150, retention rate = 0.35, paid amount = 2000], a total of 6 dimensions. This multidimensional delivery feature is then dedimensionalized, normalizing each feature value to the range of 0 to 1, resulting in the first multidimensional feature, for example [0.5, 0.4, 0.3, 0.25, 0.35, 0.4]. This first multidimensional feature is then input into a fully connected layer with 6 input neurons and 6 output neurons. The weight matrix of the fully connected layer is a 6×6 matrix, and the bias is a 6-dimensional vector. After matrix multiplication and ReLU activation function transformation, a 6-dimensional latent representation vector is output, for example [0.82, 0.15, 0.63, 0.41, 0.77, 0.29]. The first dimension 0.82 in the implicit representation vector is no longer a simple scaling of the dedimensionalized consumption amount to 0.5, but rather the result of a weighted combination of six dedimensionalized feature values: consumption amount, impressions, clicks, activations, retention rate, and payment amount, after non-linear activation. Similarly, the other dimensions also incorporate all six dedimensionalized feature information.

[0026] Through the above implementation method, the nonlinear transformation of the fully connected layer enables each dimension of the latent representation vector to incorporate the interactive information of multiple original projection features, overcoming the defects of independent processing of each feature and loss of cross-feature relationships in the prior art; at the same time, the latent representation vector maintains the same dimension as the first multidimensional feature, and the original multidimensional projection feature also maintains the same dimension as the first multidimensional feature, thus ensuring the one-to-one correspondence between the augmented feature vector and the original multidimensional projection feature in the dimension during the subsequent element-wise augmentation operation.

[0027] S12. The attention weights corresponding to each feature dimension are obtained by performing attention calculation on the latent representation vector through an attention network.

[0028] After obtaining the latent representation vector in S11, an attention network is used to calculate the attention weights for each feature dimension. The attention network is a neural network substructure used to calculate the importance of each feature dimension to the prediction target in the current sample; it can consist of at least one fully connected layer and a normalization function (such as the Softmax function). The attention weight is a scalar value output by the attention network for each feature dimension of the latent representation vector. The sum of the attention weights for all feature dimensions is 1. The larger the attention weight of a single feature dimension, the more important that feature dimension is for the return on investment prediction in the current sample. Attention calculation refers to the complete calculation process from the latent representation vector to the attention weights.

[0029] For example, the d-dimensional latent representation vector output from S11 is input into the attention network. The attention network first maps the latent representation vector into d attention scores (unnormalized real values) through a fully connected layer, with each attention score corresponding to a feature dimension. Then, the d attention scores are normalized using the Softmax function, so that the sum of the attention weights for all feature dimensions is 1, thus obtaining the attention weights corresponding to each feature dimension.

[0030] In one embodiment, following the example in S11 above, after the 6-dimensional latent representation vector [0.82, 0.15, 0.63, 0.41, 0.77, 0.29] is input into the attention network, the fully connected layer of the attention network outputs 6 attention scores, for example [2.1, -0.3, 1.5, 0.2, 1.8, -0.8]. After Softmax normalization, 6 attention weights are obtained, for example [0.28, 0.05, 0.22, 0.09, 0.25, 0.11]. Among them, the attention weight of the first dimension (the latent representation corresponding to the original consumption amount) is the highest at 0.28, indicating that this feature dimension is the most important for ROI prediction under the current sample; the attention weight of the second dimension (the latent representation corresponding to the original number of impressions) is the lowest at 0.05, indicating that this feature dimension is relatively less important for ROI prediction under the current sample. Attention weights are calculated based on latent representation vectors, which in turn incorporate interaction information from multiple original features. Therefore, attention weights do not reflect the importance of a single original feature, but rather the importance of the overall feature dimension after cross-feature interaction.

[0031] Through the above implementation method, the importance of the latent representation vector that integrates cross-feature interaction information is evaluated dimension by dimension by using the attention network. This enables the model to dynamically identify which feature dimensions are more important to the return on investment prediction based on the specific feature values ​​of each ad sample. This overcomes the defect of treating all features equally in the prior art. Since the attention weights change with different samples (for example, the feature values ​​of the same ad are different at different times, and its attention weight distribution is also different), adaptive feature importance evaluation is achieved for each sample.

[0032] S13. Obtain the value distribution of each feature dimension in multiple historical ads of the same type as the target ad. Determine the gate value corresponding to each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The feature dimension whose current value deviates further from its value distribution has a lower gate value.

[0033] After obtaining the attention weights corresponding to each feature dimension in S12, the value distribution of each feature dimension in multiple historical ads of the same type as the target ad is obtained. The gate value for each feature dimension is determined based on the deviation of its current value from its value distribution. Here, "same type of historical ad" refers to an ad that is similar to the target ad in at least one dimension, such as ad type, delivery channel, promotional content category, and target audience profile, and has already been delivered or has accumulated sufficient historical data. The value distribution refers to the statistical distribution of the values ​​of multiple similar historical ads on the same feature dimension, which can be represented by the mean and standard deviation, or by quantile intervals. The deviation refers to the degree of difference between the current feature value of the target ad and the center of the value distribution of that feature dimension in historical ads (such as the mean or median), which can be measured by how many standard deviations the current value deviates from the mean (i.e., the absolute value of the standard score). The gate value is a scalar value between 0 and 1 corresponding to each feature dimension, representing the reliability of the information of that feature dimension in the current sample. The feature dimension whose current value deviates further from its value distribution has a lower gate value. This can also be understood as follows: a higher gate value indicates that the value of the feature dimension is closer to the historical normal range and the information is more reliable; a lower gate value indicates that the value of the feature dimension deviates more from the historical normal range and is more likely to be an outlier.

[0034] For example, firstly, historical ads of the same type as the target ad are retrieved from the historical database. For each feature dimension, the historical values ​​of all similar historical ads on that feature dimension are collected, and the mean and standard deviation are calculated as the value distribution. Then, for each feature dimension in the multi-dimensional delivery features of the target ad, the deviation of its current value from the value distribution of that feature dimension is calculated. The deviation is calculated as follows: ,in, Indicates the first The target advertisement in the first Deviation scores on each feature dimension Indicates the first The target advertisement in the first The current value in each feature dimension Indicates the first The average of each feature dimension across multiple historical ads of the same type. Indicates the first The standard deviation of each feature dimension across multiple historical ads of the same type is used to calculate the deviation score. Therefore, this formula represents the deviation score obtained by subtracting the mean from the current value, taking the absolute value, and then dividing by the standard deviation. A preset mapping function then maps the deviation score to a gate value between 0 and 1, with the larger the deviation score, the lower the gate value. For example, a gate value of 1.0 is used when the deviation score is 0 (i.e., the current value equals the historical mean); a gate value of 0.3 is used when the deviation score is 2 (i.e., deviating from 2 standard deviations); and the gate value approaches 0 when the deviation score exceeds 3.

[0035] It should be noted that the gating value calculation in step S13 and the attention weight calculation in step S12 are independent of each other in terms of input information source. The attention weight in step S12 is calculated from the latent representation vector (i.e., the nonlinear interactive representation after encoding by the fully connected layer), reflecting the importance of each feature dimension to the ROI prediction. The gating value in step S13 is calculated from the deviation of the current value of the original multi-dimensional advertising features from the distribution of similar historical advertising values, without going through the fully connected layer encoding path, reflecting whether the value of each feature dimension in the current sample is abnormal and whether the information is reliable. The two calculation paths evaluate each feature dimension independently from the two dimensions of importance and reliability, and are independent of each other.

[0036] In one embodiment, following the previous example, for the feature dimension of spending amount, it is assumed that the average spending amount of similar historical advertisements is 4000 yuan, and the standard deviation is 800 yuan. The current spending amount of the target advertisement is 5000 yuan, and the deviation score = (5000-4000) / 800 = 1.25. According to the preset mapping function (such as using an exponential decay function), the gate value corresponding to the deviation score of 1.25 is approximately 0.65. For the feature dimension of user payment amount, it is assumed that the average payment amount of similar historical advertisements is 2500 yuan, and the standard deviation is 600 yuan. The current payment amount of the target advertisement is 2000 yuan, and the deviation score = |2000-2500| / 600≈0.83. The gate value corresponding to the deviation score of 0.83 is approximately 0.78.

[0037] Through the above implementation method, by determining the gate value based on the deviation of the current feature value from the distribution of similar historical ad values, automatic detection and quantitative evaluation of abnormal feature values ​​in advertising data are achieved. For example, when the current value of a certain feature dimension deviates abnormally due to external factors such as population fluctuations or market competition, its gate value is automatically reduced, and the contribution of this feature dimension in subsequent ROI prediction is automatically suppressed, thereby reducing the interference of outliers on the ROI prediction results. At the same time, the calculation path of the gate value is independent of the calculation path of the attention weight. That is, the gate value is not encoded by the fully connected layer, but is directly calculated from the original features and historical distribution. This makes the reliability evaluation of feature values ​​and the adaptive learning of feature importance two decoupled dimensions that do not interfere with each other. This overcomes the defect of limited evaluation accuracy caused by mixing two feature evaluation dimensions of different natures in the same processing path in the prior art, and thus significantly improves the accuracy of subsequent ROI prediction.

[0038] S14. Multiply the value of each dimension of the latent representation vector, the attention weight of the corresponding dimension, and the gating value of the corresponding dimension element by element to obtain the enhanced feature vector.

[0039] After obtaining the attention weights in S12 and the gating values ​​in S13, the values ​​of each dimension of the latent representation vector, the corresponding attention weights, and the corresponding gating values ​​are multiplied element-wise to obtain the enhanced feature vector. Element-wise multiplication refers to multiplying the three scalar values ​​at the same index position of the three vectors of the same dimension (i.e., the latent representation vector, the attention weight vector, and the gating value vector) to obtain a new vector of the same dimension. The enhanced feature vector is the feature vector enhanced by the attention weights and gating values. Its dimension is the same as the latent representation vector, but the value of each dimension is the result of multiplying the original value of the latent representation vector by the attention weight and gating value of that dimension.

[0040] For example, the three d-dimensional vectors—the latent representation vector, the attention weight vector, and the gating value vector—are multiplied element-wise, for example, for the dimension index. ( =1,2,…,d), the first eigenvector of the enhanced feature vector Dimension value = the first dimension of the implicit representation vector The first dimension value × attention weight Dimension value × gate value Dimensionality. Since both attention weights and gating values ​​are between 0 and 1 (or sum to 1), the element-wise multiplication results in the following: for dimensions with high attention weights (i.e., important dimensions), their latent representation vector values ​​are largely preserved; for dimensions with low attention weights (i.e., unimportant dimensions), their latent representation vector values ​​are weakened; for dimensions with low gating values ​​(i.e., dimensions with anomalous or unreliable values), even if their attention weights are high, their latent representation vector values ​​are still weakened by the gating value. Only those dimensions that are both deemed important by the attention network and reliable by the gating mechanism have their contributions to the enhanced feature vector fully preserved.

[0041] In one embodiment, following the foregoing example, assume the latent representation vector is [0.82, 0.15, 0.63, 0.41, 0.77, 0.29], the attention weights are [0.28, 0.05, 0.22, 0.09, 0.25, 0.11], and the gating value is [0.65, 0.90, 0.78, 0.92, 0.70, 0.78]. After element-wise multiplication, the enhanced feature vector is: [0.82×0.28×0.65,0.15×0.05×0.90,0.63×0.22×0.78,0.41×0.09×0.92,0.77×0.25×0.70,0.29×0.11×0.78] = [0.149,0.007,0.108,0.034,0.135,0.025]. As can be seen, the second dimension (the dimension corresponding to the original number of impressions) was significantly suppressed due to its extremely low attention weight (0.05); although the first dimension (the dimension corresponding to the original consumption amount) had the highest attention weight (0.28), its enhanced contribution of 0.149 was suppressed by about 35% compared to the result of enhancing only the attention weight (0.82×0.28=0.230) because the gating value was only 0.65 (the current consumption amount deviates from the historical average by 1.25 standard deviations).

[0042] Through the above implementation method, by combining the element-wise effects of attention weights and gating values, both the importance of the dimension to the return on investment prediction in the current sample (characterized by attention weights) and the reliability of the dimension's value in the current sample (characterized by gating values) are considered. This achieves a dual screening of each feature dimension in the latent representation vector, ensuring that only feature dimensions that are both important and reliable contribute significantly to the enhanced feature vector. This overcomes the prediction bias caused by the complete retention of important but anomalous features in existing technologies, as well as the efficiency reduction caused by reliable but unimportant features occupying model capacity. As a result, prediction accuracy and prediction efficiency are significantly improved.

[0043] S15. Concatenate the multi-dimensional delivery features and the enhanced feature vector along the feature dimension to obtain the fused feature set, and perform the first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment of the target advertisement.

[0044] After obtaining the enhanced feature vector in S14, the multi-dimensional delivery feature and the enhanced feature vector are concatenated along the feature dimension to obtain a fused feature set. This fused feature set is then subjected to first gradient boosting decision processing to obtain the estimated ROI of the target advertisement. Concatenating along the feature dimension means connecting two vectors with the same or different dimensions along the feature dimension to form a new vector with a dimension equal to the sum of the two dimensions. First gradient boosting decision processing involves predicting the input features using a gradient boosting decision tree model. This model is an ensemble learning model based on decision trees, trained sequentially with multiple decision trees. Each new tree fits the prediction residuals of its predecessor trees, and the prediction results from multiple trees are summed to obtain the final prediction result. The estimated ROI is the value output by the gradient boosting decision tree model, representing the ratio of future revenue to investment for the target advertisement.

[0045] For example, the d-dimensional enhanced feature vector obtained in S14 is concatenated with the d-dimensional original multi-dimensional delivery feature obtained in S11 along the feature dimension. For instance, the d values ​​of the enhanced feature vector are sequentially arranged after the d values ​​of the original multi-dimensional delivery feature to form a 2d-dimensional fused feature set. Then, the fused feature set is input into a pre-trained gradient boosting decision tree model. The gradient boosting decision tree model contains M decision trees, where M is a preset hyperparameter, such as 100 trees. Each decision tree makes an independent path decision on the fused feature set and outputs a predicted value. The predicted value of the first tree is directly output. Starting from the second tree, each tree fits the residual between the predicted value of the preceding tree and the true value. Finally, the predicted values ​​of the M trees are summed to obtain the estimated return on investment (ROI) of the target advertisement.

[0046] It should be noted that the fused feature set includes both enhanced feature vectors that have undergone attention enhancement and gating filtering, and multi-dimensional projection features that retain the original value distribution. Since the enhanced feature vectors, although subjected to noise suppression and importance weighting, may lose some distribution information from the original features, concatenating the original and enhanced features and feeding them simultaneously into the gradient boosting decision tree allows the tree model to autonomously choose whether to use the original or enhanced features, or a combination of both, during the splitting process. This achieves both the high signal-to-noise ratio advantage of enhanced features and the complete information retention advantage of the original features.

[0047] In one embodiment, following the previous example, the original multidimensional deployment features are [5000, 100000, 3000, 150, 0.35, 2000] (6 dimensions), and the enhanced feature vector is [0.149, 0.007, 0.108, 0.034, 0.135, 0.025] (6 dimensions). After concatenation along the feature dimensions, a 12-dimensional fused feature set is obtained: [5000, 100000, 3000, 150, 0.35, 2000, 0.149, 0.007, 0.108, 0.034, 0.135, 0.025]. This fused feature set is input into a gradient boosting decision tree model containing 100 decision trees. The model processes each tree and then sums the results to output the estimated return on investment, for example, 1.42.

[0048] Through the above implementation method, by concatenating the enhanced feature vector with the original multidimensional delivery features, the gradient boosting decision tree can simultaneously utilize high-quality enhanced features and complete original features. When splitting nodes, the tree model can flexibly choose the original feature dimension or the enhanced feature dimension for decision-making, achieving complementarity between enhanced and original information. In addition, by using the gradient boosting decision tree as the prediction model, and leveraging its powerful nonlinear fitting ability and natural adaptability to tabular data, it can more accurately capture the complex nonlinear relationship between multidimensional delivery features and return on investment compared to a single linear model, thereby significantly improving the accuracy of return on investment prediction.

[0049] S16. Perform placement control operations on the target advertisement based on the estimated return on investment.

[0050] After obtaining the estimated ROI in S15, the target ad is subject to performance control operations based on this ROI. Performance control operations refer to adjustments made to the target ad's performance status based on the estimated ROI, including continuing performance, reducing the budget, and shutting down. Continuing performance means maintaining the target ad's current budget and bidding strategy, continuing bidding on the advertising platform. Reducing the budget means lowering the target ad's daily or total budget cap to reduce subsequent costs. Shutting down means ceasing performance of the target ad and ceasing bidding for display.

[0051] For example, the estimated ROI obtained in S15 is compared with a preset decision threshold, and the appropriate delivery control operation is determined based on the comparison result. The decision threshold includes an upper limit threshold and a lower limit threshold for the ROI. For example, the upper limit threshold is 1.2 and the lower limit threshold is 0.8, and the specific values ​​of both can be configured according to the actual situation. When the estimated ROI is greater than or equal to the upper limit threshold, a continue delivery instruction is sent to the delivery management interface of the advertising platform; when the estimated ROI is less than the upper limit threshold but greater than or equal to the lower limit threshold, a budget reduction instruction is sent to the advertising platform to reduce the daily budget of the target advertisement to a preset percentage of the current budget (e.g., 50%); when the estimated ROI is less than the lower limit threshold, a shutdown instruction is sent to the advertising platform to stop the bidding delivery of the target advertisement.

[0052] In one embodiment, following the previous example, the estimated ROI of the target ad is 1.42. Assume the upper threshold for ROI is 1.2 and the lower threshold is 0.8. Since 1.42 is greater than 1.2, it is determined that a continue-delivery operation will be performed, and an instruction to maintain the current delivery configuration of the target ad (i.e., a continue-delivery instruction) is sent to the advertising platform.

[0053] Through the above implementation methods, by directly using the estimated return on investment (ROI) to drive the ad placement control operation, a fully automated closed loop is achieved from feature collection to feature enhancement to ROI estimation and then to ad placement decision-making. This avoids the delays and subjective biases caused by manual data interpretation and manual operation in existing technologies, thereby improving the response speed of ad placement decisions from the hourly level to the real-time level. It also avoids budget waste caused by decision delays, thus significantly improving the objectivity and timeliness of ad placement control operations.

[0054] As described above, a latent representation vector incorporating cross-feature interaction information is obtained through nonlinear transformation of a fully connected layer. An attention network learns the sample-by-sample importance weights of each feature dimension from the latent representation vector. An independent gating path based on the deviation of historical value distribution is used to evaluate the reliability of each feature dimension's values. Then, the attention weights and gating values ​​are combined and applied to the latent representation vector through element-wise multiplication to obtain an enhanced feature vector. This strengthens feature dimensions that are deemed important by both attention and reliable by gating, while suppressing unimportant or unreliable feature dimensions. Simultaneously, the enhanced feature vector is concatenated with the original multi-dimensional delivery features and fed into a gradient boosting decision tree for ROI prediction (i.e., the first gradient boosting decision processing), enabling the tree model to utilize both high-quality enhanced features and complete original features. Finally, delivery control operations are directly executed based on the predicted ROI, achieving an automated closed loop from feature acquisition to decision execution. This solves the technical problems of low prediction accuracy and decision delays caused by manual operation in existing technologies, significantly improving the accuracy of ROI estimation and consequently enhancing the accuracy and timeliness of advertising delivery operations.

[0055] The above describes a process where, during ad delivery, multi-dimensional delivery features of the target ad are acquired, and these features are dedimensionalized to obtain a first multi-dimensional feature. This first multi-dimensional feature is then input into at least one fully connected layer for nonlinear transformation to obtain a latent representation vector. An attention network is used to perform attention calculations on the latent representation vector to obtain the attention weights corresponding to each feature dimension. The value distribution of each feature dimension in multiple historical ads of the same type as the target ad is obtained, and the gate value corresponding to each feature dimension is determined based on the deviation of the current value of each feature dimension from its value distribution. The value of each dimension in the latent representation vector, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension are multiplied element-wise to obtain an enhanced feature vector. The multi-dimensional delivery features and the enhanced feature vector are concatenated along the feature dimensions to obtain a fused feature set, and the fused feature set is subjected to a first gradient boosting decision process to obtain the estimated return on investment (ROI) of the target ad. The delivery control operation for the target ad is then performed based on the estimated ROI. By employing the aforementioned techniques, attention weights are used to learn which features are more important for ROI prediction from the latent representation of feature interactions. Gating values ​​are used to assess which features are abnormal or unreliable in the current sample based on the deviation of the original values ​​from the historical distribution. These two methods, starting from different information sources and following independent computational paths, measure the importance and reliability of features respectively. Then, element-wise multiplication is used to jointly apply the enhanced feature vector to the latent representation vector, thus strengthening important and reliable features while suppressing unimportant or unreliable features. Simultaneously, the enhanced feature vector is concatenated with the original multidimensional delivery features, enabling the first gradient boosting decision processing to utilize both the enhanced high-quality features and the original features that retain the original distribution information. This addresses the technical problem in existing technologies where all features are treated equally and noisy features also participate in the estimation, leading to low ROI prediction accuracy. This embodiment significantly improves the accuracy of ROI prediction by outputting the corresponding predicted ROI through the first gradient boosting decision processing. Subsequently, delivery control operations are performed on the target advertisement based on the predicted ROI, thereby improving the accuracy of the advertisement delivery operation.

[0056] Figure 2 This is a flowchart of another advertising delivery method provided in the embodiments of this application, see below. Figure 2 When determining the latent representation vector, this ad delivery method specifically includes: S111. Identify the data type of each feature in the multidimensional delivery features, keep the original values ​​of numerical features unchanged, and map categorical features to numerical scalars through the embedding layer to obtain a standardized feature set.

[0057] After obtaining the multi-dimensional delivery features of the target advertisement in S11, the data type of each feature in the multi-dimensional delivery features is identified. Numerical features retain their original values, while categorical features are mapped to numerical scalars through an embedding layer, resulting in a standardized feature set. Here, data type refers to the data attributes of the feature columns, including numerical and categorical features. Numerical features are indicators that are themselves continuous values, such as spending amount (5000 yuan) or impressions (100,000 times), and their numerical magnitude has direct mathematical meaning. Categorical features are indicators that take values ​​of a finite number of discrete categories, such as ad placement type (values ​​can be "splash screen ad," "feed ad," or "rewarded video ad," etc.) or operating system type; the numerical relationships between categories do not have direct mathematical meaning. The embedding layer is a neural network layer that maps categorical features to dense numerical vectors. Its core is a learnable embedding matrix, where the number of rows in the matrix represents the total number of value categories for that categorical feature, and the number of columns represents the preset embedding dimension. A categorical value is converted into a numerical scalar through table lookup and vector compression operations. A standardized feature set refers to a feature set in which all feature columns have been converted into a uniform numerical format (i.e., each feature column is a scalar value).

[0058] For example, after obtaining the multidimensional targeting features of the target ad, each feature column in the multidimensional targeting features is traversed to determine its data type. For numerical feature columns, no transformation is performed, and their original values ​​remain unchanged. For categorical feature columns, an embedding lookup operation is first performed on the categorical value through an embedding layer. For example, the corresponding embedding vector (a dense floating-point vector) is found in the embedding matrix using the categorical value as the row index. Then, the embedding vector is compressed into a numerical scalar through operations such as averaging or linear projection. After processing all feature columns, a standardized feature set containing only numerical scalars is obtained. The standardized feature set has the same dimensions as the original multidimensional targeting features, that is, each feature column still corresponds to a scalar value, but all columns are in numerical scalar format, which can be uniformly normalized before being input into a fully connected layer for processing.

[0059] In one embodiment, assume the multi-dimensional delivery feature contains 7 feature columns, of which 5 columns are numerical, such as spending amount = 5000, impressions = 100000, clicks = 3000, activations = 150, and retention rate = 0.35; 2 columns are categorical, such as ad placement type "feed ad" and operating system type "Type A". For the category value "feed ad", the embedding layer first looks up an embedding vector in the embedding matrix, for example [0.32, -0.15, 0.48, 0.21] (embedding dimension set to 4), and then averages the 4 values ​​of the embedding vector to obtain a numerical scalar of 0.215. Similarly, "Type A" is embedded and compressed to obtain a numerical scalar of 0.67. The final standardized feature set is [5000, 100000, 3000, 150, 0.35, 0.215, 0.67], a total of 7 numerical scalars.

[0060] Through the above implementation methods, by recognizing data types and embedding categorical features, the use of the embedding layer enables the semantic relationships between the values ​​of categorical features to be learned and preserved. For example, "splash screen ads" and "feed ads" may be mapped to relatively close positions in the embedding space because they both belong to display ad slots. Compared with the traditional one-hot encoding method, the embedding mapping not only significantly compresses the feature dimension and avoids dimensionality explosion when there are many categories, but also provides continuous numerical inputs containing the semantic relationships between categories for subsequent normalization processing and fully connected layers.

[0061] S112. Normalize the standardized feature set to obtain the first multidimensional feature.

[0062] After obtaining the standardized feature set in S111, the standardized feature set is normalized to obtain the first multidimensional feature. For example, the d-dimensional standardized feature set obtained in S111 (where all feature columns are numerical scalars) is normalized, mapping each feature value to a uniform numerical range to obtain the d-dimensional first multidimensional feature. This first multidimensional feature is then input into at least one pre-constructed fully connected layer. The fully connected layer processes this step in the same way as in S11, transforming the first multidimensional feature into a d-dimensional latent representation vector through a linear combination of weight matrices and nonlinear activation functions such as ReLU. Since S111 has already converted categorical features into numerical scalars, and this step has eliminated the differences in dimensions and magnitudes between features, the input received by the fully connected layer is a uniform continuous numerical format, allowing for normal matrix multiplication and nonlinear transformations.

[0063] In one embodiment, following the example of S111 above, the 7-dimensional standardized feature set [5000, 100000, 3000, 150, 0.35, 0.215, 0.67] is normalized to obtain a 7-dimensional first multi-dimensional feature, such as [0.5, 0.4, 0.3, 0.25, 0.35, 0.3, 0.5]. This first multi-dimensional feature is input into a fully connected layer (7×7 weight matrix), and after nonlinear transformation, a 7-dimensional latent representation vector is obtained, such as [0.74, 0.18, 0.56, 0.39, 0.82, 0.45, 0.31].

[0064] Through the above implementation method, the mixed-type features are first unified into a numerical format through data type identification and embedding mapping. Then, the differences in dimensions and magnitudes between features are eliminated through normalization processing, so that the fully connected layer can correctly process the mixed-type features. The cross-type interaction relationship between numerical features and categorical features is captured through nonlinear transformation. For example, the nonlinear combination between consumption amount and ad slot type affects the return on investment. This avoids the loss of feature information caused by inconsistent types or dimensions, thereby improving the accuracy of the obtained latent representation vector and thus improving the accuracy of the subsequent estimated return on investment.

[0065] As described above, by identifying and processing numerical and categorical features separately, the categorical features are mapped to numerical scalars through the embedding layer while retaining their semantic relationships between categories. Compared with traditional one-hot encoding, the embedding mapping provided in this implementation significantly compresses the feature dimension, avoids the dimension explosion caused by high cardinality categorical features, and retains the semantic information between categories, thus providing a unified input format for subsequent normalization processing and correct reception and processing by the fully connected layer.

[0066] Figure 3 This is a flowchart of another advertising delivery method provided in the embodiments of this application, referred to... Figure 3 When determining the estimated return on investment for a target advertisement, this advertising placement method specifically includes: S21. Concatenate the multidimensional projection features and the enhanced feature vector along the feature dimension to obtain the fused feature set.

[0067] Following S14, the multidimensional projection features and enhanced feature vectors are concatenated along the feature dimensions to obtain a fused feature set. The processing procedure in this step is the same as the concatenation operation in S15, and will not be repeated here.

[0068] S22. In order of time nodes from near to far, perform first gradient boosting decision processing on multiple preset time nodes in sequence to obtain the estimated output ratio corresponding to each preset time node; wherein, the processing of the first preset time node takes the fused feature set as input, and the processing of each subsequent preset time node takes the feature set obtained by splicing the fused feature set with the estimated output ratio of the previously processed preset time node as input, so that the estimated output ratio of the subsequent preset time node is based on the estimated results of the previous preset time node.

[0069] After obtaining the fused feature set in S21, the first gradient boosting decision processing is performed sequentially on multiple preset time nodes in ascending order of time nodes to obtain the estimated output-return ratio for each preset time node. Here, the preset time node refers to a pre-defined future time point used to predict the output-return ratio, such as 24 hours, 72 hours, and 168 hours from the current time. The sequential condition processing means that when estimating multiple time nodes sequentially, the prediction of subsequent time nodes is not only based on the fused feature set but also additionally receives the estimated output-return ratio of the previously predicted time nodes as additional input conditions, forming a prediction chain.

[0070] For example, after obtaining the fused feature set in S21, several preset time nodes that need to be estimated are first determined, and these time nodes are arranged in order of proximity to the current time. For the most recent time node (i.e., the first preset time node), the fused feature set obtained in S21 is directly used as the input to the gradient boosting decision tree model, and the estimated return on investment for that time node is output. For the second time node, the fused feature set is concatenated with the estimated return on investment for the first time node, that is, an additional dimension (the estimated value of the first node) is added to the fused feature set. The concatenated feature set is used as the input to the gradient boosting decision tree model, and the estimated return on investment for the second time node is output. This process proceeds sequentially, so that the estimation of each time node is based on the estimation result of the previous time node, forming a prediction chain from near to far.

[0071] In one embodiment, it is assumed that the ROI needs to be estimated for three time points: the first day (T+1), the third day (T+3), and the seventh day (T+7). The fused feature set is a 12-dimensional vector. First, using the 12-dimensional fused feature set as input, the gradient boosting decision tree outputs the estimated ROI for the first day as 1.15. Then, the 12-dimensional fused feature set is concatenated with the estimated ROI for the first day (1.15) to obtain a 13-dimensional vector, which is then used as input, and the gradient boosting decision tree outputs the estimated ROI for the third day as 1.32. Finally, the 12-dimensional fused feature set is concatenated with the estimated ROI for the third day (1.32) to obtain a 13-dimensional vector, which is then used as input, and the gradient boosting decision tree outputs the estimated ROI for the seventh day as 1.48.

[0072] Through the above implementation method, by using the estimated results of the preceding time nodes as additional input conditions for the estimated results of the subsequent time nodes, the estimated results of the subsequent time nodes can utilize the estimated information of the preceding time nodes. For example, if the estimated production ratio on the first day is high, the estimated production ratio for the three days may be further revised upward based on the first day, forming an information transmission chain across time nodes. This ensures that the estimates of each time node are not isolated from each other but sequentially related. Compared with the existing method of independently estimating each time node, this embodiment can more accurately capture the trend of the production ratio evolving over time, thereby improving the accuracy of subsequent control operations based on the estimated production ratio.

[0073] S23. Calculate the confidence level of the trend direction based on the consistency of the change direction of the estimated production ratio between adjacent preset time nodes. The confidence level of the trend direction is the proportion of the number of intervals where adjacent preset time nodes change in the same direction to the total number of intervals.

[0074] After obtaining the estimated production ratio for each preset time node in S22, the confidence level of the trend direction is calculated based on the consistency of the direction of change of the estimated production ratio between adjacent preset time nodes. Adjacent preset time nodes refer to two adjacent nodes in a sequence of preset time nodes arranged from nearest to farthest in time, such as the first day node and the third day node, or the third day node and the seventh day node, each being an adjacent pair. The direction of change refers to the direction of increase or decrease of the estimated production ratio between adjacent nodes. For example, if the estimated production ratio of a subsequent node is greater than that of an preceding node, the direction of change is upward; if the estimated production ratio of a subsequent node is less than that of an preceding node, the direction of change is downward. The confidence level of the trend direction is used to quantify the consistency and reliability of the overall trend direction. Its value is the proportion of the number of intervals in all pairs of adjacent nodes whose direction of change is consistent with the overall trend direction to the total number of intervals.

[0075] For example, after obtaining the estimated production ratio corresponding to each preset time node in S22, the direction of change of the estimated production ratio for each pair of adjacent preset time nodes is first calculated. Then, the overall trend direction is statistically analyzed. For example, if most adjacent node pairs show an upward trend, the overall trend direction is determined to be upward. Finally, the confidence level of the trend direction is calculated, which is the number of adjacent node pairs consistent with the overall trend direction divided by the total number of adjacent node pairs.

[0076] In one embodiment, following the example of S22 above, the estimated production ratios for the three time points are 1.15 (first day), 1.32 (third day), and 1.48 (seventh day). There are two pairs of adjacent nodes: one pair shows a change of +0.17 from the first day to the third day, with an upward trend; the other pair shows a change of +0.16 from the third day to the seventh day, also with an upward trend. Both pairs show an upward trend, the overall trend is upward, and all adjacent pairs are consistent with the overall trend direction. Therefore, the confidence level for the trend direction is 2 / 2 = 1.0 (i.e., 100%). As another example, if the estimated production ratios are 1.15, 1.08, and 1.42, then the adjacent changes are -0.07 from the first day to the third day, with a downward trend; and +0.34 from the third day to the seventh day, with an upward trend. The two pairs of changes have inconsistent directions, the overall trend direction is uncertain, and the confidence level is low.

[0077] By introducing the quantitative indicator of trend direction confidence through the above implementation method, the deployment control operation relies not only on the absolute value of the estimated return ratio but also on the consistency of the trend direction. For example, a consistent trend (high confidence) is more reliable, while a fluctuating trend (low confidence) reflects potential instability in the estimated result. This avoids making aggressive operational decisions when the trend direction is unstable, reduces the risk of misoperation due to accidental fluctuations, and ultimately improves the accuracy of the deployment control operation finally determined based on the trend direction confidence.

[0078] S24. Determine the deployment control operation based on the trend direction of the time-series trajectory formed by the confidence level of the trend direction and the estimated production ratio of each preset time node.

[0079] After obtaining the trend direction confidence level in S23, the deployment control operation is determined based on the trend direction of the time-series trajectory formed by the trend direction confidence level and the estimated return on investment at each preset time node. Here, the time-series trajectory refers to a broken line or point sequence formed by connecting the estimated return on investment at each time node in chronological order, with time as the horizontal axis and the estimated return on investment at each time node as the vertical axis. In one embodiment, the deployment control operation is determined according to preset decision rules by combining the trend direction confidence level calculated in S23 and the overall trend direction of the time-series trajectory. The following S25-S29 are specific elaborations of the decision rules in S24 under different scenarios.

[0080] S25. When the time sequence trajectory shows an upward trend and the estimated output ratio of the most recent preset time node is greater than or equal to the first preset threshold and the confidence level of the trend direction is greater than or equal to the preset direction confidence threshold, the continued deployment operation is executed.

[0081] Following the trend of the time-series trajectory formed in S24, if the time-series trajectory shows an upward trend and the estimated return on investment (ROI) of the most recent preset time node is greater than or equal to the first preset threshold, and the confidence level of the trend direction is greater than or equal to the preset direction confidence threshold, then the continued deployment operation is executed. For example, assuming the estimated ROI for the first day, third day, and seventh day are 1.15, 1.32, and 1.48 respectively, showing an upward trend, the estimated ROI of the most recent node (i.e., the seventh day) of 1.48 is greater than or equal to the first preset threshold of 1.2, and the confidence level of the trend direction of 1.0 is greater than or equal to the preset direction confidence threshold of 0.7, then the continued deployment operation is executed.

[0082] Through the above implementation methods, when the return on investment trajectory is consistently upward and the final node has reached the profitability line and the directional confidence is high, it indicates that the future monetization ability of the advertisement is continuously enhanced. At this time, continuing to invest can make full use of the upward momentum of the advertisement to obtain more revenue, and avoid prematurely shutting down the potential advertisement due to fluctuations at a single point in time.

[0083] S26. When the time sequence trajectory shows an upward trend but the estimated output ratio of the most recent preset time node is less than the first preset threshold, a budget reduction operation is performed.

[0084] Following the trend of the time-series trajectory formed in S24, if the time-series trajectory shows an upward trend but the estimated return on investment at the nearest preset time node is less than the first preset threshold, a budget reduction operation is executed. For example, suppose the estimated return on investment for the first day, three days, and seven days are 0.65, 0.78, and 0.95, respectively, showing an upward trend. However, the estimated return on investment for the nearest node (i.e., seven days) is 0.95, which is less than the first preset threshold of 1.2. Although the trend is positive, it has not yet reached the profitability line. Therefore, a budget reduction operation is executed to reduce the consumption rate and extend the observation time.

[0085] Through the above implementation methods, for advertisements that are showing a continuous improvement trend but have not yet become profitable, the budget is reduced instead of being shut down directly. This allows for sufficient optimization time while controlling risks, and avoids mistakenly killing potential advertisements due to short-term failure to reach profitability.

[0086] S27. When the time series trajectory shows an upward trend and the estimated output ratio of the most recent preset time node is greater than or equal to the first preset threshold, but the confidence level of the trend direction is less than the preset direction confidence threshold, the budget reduction operation is executed and the data observation period is extended.

[0087] Following the trend of the time-series trajectory formed in S24, if the time-series trajectory shows an upward trend and the estimated return on investment at the nearest preset time node is greater than or equal to the first preset threshold, but the confidence level of the trend direction is less than the preset direction confidence level threshold, a budget reduction operation is performed and the data observation period is extended. For example, assuming the estimated return on investment for the first day, three days, and seven days are 0.85, 1.45, and 1.30 respectively, although the overall trend is upward and the estimated return on investment at the seven-day node (1.30) is greater than or equal to the first preset threshold (1.2), the adjacent change directions are: from the first day to the third day, the change is +0.60 (upward); from the third day to the seventh day, the change is -0.15 (downward). These two pairs of directions are inconsistent, and the trend direction confidence level = 1 / 2 = 0.5, which is less than 0.7. Although the target is met on the surface, there are fluctuations in the direction, possibly due to data volatility. Therefore, a budget reduction operation is performed and the observation period is extended to await further data confirmation.

[0088] Through the above implementation method, false optimistic signals that meet the target value but are unstable in direction are filtered out by the confidence level of the trend direction. That is, although the return on investment of the advertisement has reached the profitability line at the most recent point, the repeated trend indicates that the forecast itself may not be stable enough. At this time, instead of continuing to increase the volume, the budget is reduced and more data is waited for, thus avoiding making aggressive decisions based on accidental high values ​​when the forecast is unstable.

[0089] S28. When the time sequence trajectory shows a downward trend or a stable trend and the estimated production ratio of the most recent preset time node is less than the second preset threshold, a shutdown operation is performed.

[0090] Following the trend of the time-series trajectory formed in S24, if the time-series trajectory shows a downward or stable trend and the estimated return on investment (ROI) at the most recent preset time point is less than the second preset threshold, a shutdown operation is performed. For example, assuming the estimated ROI for the first day, third day, and seventh day are 0.72, 0.58, and 0.41 respectively, showing a continuous downward trend, and the estimated ROI of the most recent point (i.e., the seventh day) is 0.41, which is less than the second preset threshold of 0.5, a shutdown operation is performed to stop losses in a timely manner. Through the above implementation method, for advertisements with a continuously declining ROI that have fallen below the loss threshold, a rapid shutdown is performed to stop losses.

[0091] S29. When the time sequence trajectory shows a downward trend or a stable trend and the estimated production ratio of the most recent preset time node is greater than or equal to the second preset threshold, a budget reduction operation is performed.

[0092] Following the trend of the time-series trajectory formed in S24, if the time-series trajectory shows a downward or stable trend and the estimated return on investment (ROI) at the most recent preset time node is greater than or equal to the second preset threshold, a budget reduction operation is performed. For example, assuming the estimated ROIs for the first, third, and seventh days are 1.05, 0.88, and 0.72 respectively, showing a downward trend, but the estimated ROI of the most recent node (seven days) of 0.72 is greater than or equal to the second preset threshold of 0.5, not falling below the loss threshold, and still having some recovery potential, a budget reduction operation is performed to control risk, rather than directly shutting down the advertising. Through the above implementation method, for advertisements with a declining ROI but not yet falling below the bottom line, risk and return are balanced by reducing the budget instead of immediately shutting down, avoiding premature shutdown of advertisements that still have some recovery potential.

[0093] The above-mentioned method achieves cross-time node information transmission through sequential condition prediction, distinguishes between stable and recurring trends by trend direction confidence, and combines time-series trajectory trend direction and confidence to make multi-factor joint decisions. It comprehensively covers various practical scenarios of rising, falling, or stable trends combined with whether the target is met or not, as well as high or low confidence. It solves the technical problem in existing technologies where single-point prediction cannot reflect trend changes or accidental fluctuations, leading to incorrect decisions, and achieves more robust and accurate deployment control.

[0094] The above-mentioned method achieves cross-time node information transmission through sequential condition prediction, that is, the subsequent node is based on the prediction result of the previous node, and solves the technical problem of inaccurate advertising placement operation in the prior art by quantifying the confidence of the trend direction and filtering unstable trends, and by making joint decisions based on three factors of trend direction, threshold and confidence level covering five scenarios. In the scenario of upward trend but low confidence, this embodiment can avoid making radical decisions when the information is insufficient by reducing the budget and extending the observation period, thus significantly improving the robustness and accuracy of advertising placement control.

[0095] Figure 4 This is a flowchart of another advertising delivery method provided in the embodiments of this application, see below. Figure 4 When performing ad delivery control operations, this ad delivery method specifically includes: S161. Determine the current stage of the target advertisement, wherein the current stage is determined based on the ratio of the cumulative spending budget of the target advertisement to the preset total budget; when the spending ratio is lower than the first spending ratio threshold, the current stage is determined to be the first stage; when the spending ratio is between the first spending ratio threshold and the second spending ratio threshold, the current stage is determined to be the second stage; when the spending ratio is higher than the second spending ratio threshold, the current stage is determined to be the third stage.

[0096] After obtaining the estimated ROI in S15, the first step in executing the campaign control operation is to determine the current campaign stage of the target ad. The campaign stage refers to the different phases divided according to the budget consumption progress throughout the entire lifecycle of the ad, from launch to budget exhaustion. The current campaign stage is determined based on the percentage of the target ad's cumulative budget expenditure relative to the preset total budget. The cumulative budget expenditure refers to the total amount of campaign spending actually incurred by the target ad from launch to the current moment. The preset total budget refers to the upper limit of the total budget pre-set by the advertiser for this target ad. Consumption percentage = Cumulative budget expenditure / Preset total budget.

[0097] Specifically, when the consumption ratio is lower than the first consumption ratio threshold, the current campaign stage is determined as the first stage. The first stage (which can be called the exploration stage) corresponds to the stage where the consumption ratio is lower than the first consumption ratio threshold. At this time, the ad has been online for a short time, the amount of data accumulated is limited, and the model's understanding of the ad's effectiveness is not yet sufficient.

[0098] Specifically, when the consumption ratio falls between the first consumption ratio threshold and the second consumption ratio threshold, the current campaign stage is designated as the second stage. The second stage (which can be called the optimization stage) corresponds to the stage where the consumption ratio falls between the first consumption ratio threshold and the second consumption ratio threshold. At this point, a certain amount of data has been accumulated, and the model's perception of advertising effectiveness tends to stabilize.

[0099] Specifically, when the consumption ratio exceeds the second consumption ratio threshold, the current campaign stage is designated as the third stage. The third stage (which can be called the convergence stage) corresponds to the stage where the consumption ratio exceeds the second consumption ratio threshold. At this point, the budget has been largely consumed, and a decision to continue or shut down the campaign needs to be made as soon as possible.

[0100] For example, after obtaining the estimated return on investment in S15, when performing the delivery control operation, the preset total budget and cumulative consumption budget of the target advertisement are first obtained. The consumption ratio of the target advertisement is calculated based on these two budgets. The current delivery stage of the target advertisement is determined by comparing the consumption ratio with preset first and second consumption ratio thresholds.

[0101] In one embodiment, assuming a target advertisement has a preset total budget of 50,000 yuan, and 12,000 yuan has been spent so far, the consumption rate is 12,000 / 50,000 = 0.24 (24%). The first consumption rate threshold is 20%, and the second consumption rate threshold is 60%. Since 24% is between 20% and 60%, the current stage of the target advertisement is determined to be the second stage.

[0102] Through the above implementation methods, the campaign phases are divided based on budget consumption ratios, enabling quantitative phased management of the entire advertising lifecycle and providing a clear basis for subsequent adaptive threshold adjustments.

[0103] S162. Determine the corresponding upper and lower thresholds for the return on investment (ROI) based on the current deployment stage. The upper and lower thresholds for the ROI in the first stage are lower than the corresponding thresholds in the second stage, and the upper and lower thresholds for the ROI in the third stage are higher than the corresponding thresholds in the second stage.

[0104] After determining the current campaign stage in S161, the corresponding upper and lower ROI thresholds are determined based on this stage. These upper and lower ROI thresholds are the ROI decision thresholds used to determine whether to continue or shut down the campaign, with different thresholds applied to different campaign stages. The first stage has the most lenient thresholds, with both the lower and upper ROI thresholds being relatively low, allowing ample time for data accumulation and model learning. The second stage uses standard thresholds. The third stage uses strict thresholds, forcing a final decision on whether to retain or terminate the campaign. Therefore, the upper and lower ROI thresholds for the first stage are lower than their corresponding thresholds for the second stage, while the upper and lower ROI thresholds for the third stage are higher than their corresponding thresholds for the second stage.

[0105] For example, assume that the upper limit of the ROI threshold for the first stage is 1.0 and the lower limit is 0.4; the upper limit of the ROI threshold for the second stage is 1.2 and the lower limit is 0.6; and the upper limit of the ROI threshold for the third stage is 1.5 and the lower limit is 0.9. After determining the current campaign stage in S161, assuming the current campaign stage is the second stage, the upper limit of the ROI threshold corresponding to the current campaign stage of the target ad is determined to be 1.2 and the lower limit is 0.6.

[0106] By adaptively adjusting the ROI threshold during the deployment phase, the above implementation method resolves the contradictions in fixed-threshold schemes where early-stage ads are prematurely shut down due to insufficient data and long-term ads continue to consume budgets without being stopped due to overly lenient thresholds. In this embodiment, a lower upper limit threshold for ROI is set in the first stage (exploration stage), allowing early-stage ads with potential but whose estimated ROI has not yet reached the target to accumulate more time and budget space. In the third stage (convergence stage), a higher lower limit threshold for ROI is set to ensure a decisive shutdown decision when the budget is about to run out, preventing ineffective consumption.

[0107] S163. When the estimated return on investment is greater than or equal to the upper limit threshold of the return on investment corresponding to the current stage of the campaign, continue the campaign for the target ad.

[0108] After determining the upper and lower limits of the ROI for the current campaign stage in S162, if the estimated ROI is greater than or equal to the upper limit of the ROI for the current campaign stage, the campaign continues for the target ad. The specific implementation method for continuing the campaign for the target ad is the same as that in S16, and will not be repeated here.

[0109] S164. When the estimated return on investment (ROI) is less than the upper limit threshold of ROI corresponding to the current campaign stage but greater than or equal to the lower limit threshold of ROI corresponding to the current campaign stage, a budget reduction operation is performed on the target ad.

[0110] After determining the upper and lower limits of the ROI for the current campaign stage in S162, when the estimated ROI is less than the upper limit of the ROI for the current campaign stage but greater than or equal to the lower limit of the ROI for the current campaign stage, a budget reduction operation is performed on the target ad. The specific implementation method for performing the budget reduction operation on the target ad is the same as the implementation method in S16 above, and will not be repeated here.

[0111] S165. When the estimated return on investment is less than the lower limit threshold of the return on investment for the current campaign stage, the target ad shall be shut down.

[0112] After determining the upper and lower limits of the ROI for the current campaign stage in S162, when the estimated ROI is less than the lower limit of the ROI for the current campaign stage, the target ad is shut down. The specific implementation method for shutting down the target ad is the same as that in S16, and will not be repeated here.

[0113] As described above, by introducing a campaign phase division based on budget consumption ratio and a phase-adaptive ROI threshold, the contradiction in existing technologies where fixed thresholds cannot adapt to different lifecycle stages of advertising is resolved. This embodiment is more lenient in the first stage (exploration stage) to avoid prematurely shutting down potentially promising ads; more stringent in the third stage (convergence stage) to force clear decisions on whether to retain or dismiss ads; and uses a standard threshold in the second stage (optimization stage) to balance exploration and convergence. The thresholds of the three stages transition smoothly, achieving refined campaign management throughout the entire advertising lifecycle.

[0114] Figure 5 This is a flowchart of another advertising delivery method provided in the embodiments of this application, referred to... Figure 5 The specific methods for placing advertisements include: S31. When the target ad is in the cold start stage with sparse data, the multi-dimensional delivery features are processed in a second gradient to improve decision-making and obtain an auxiliary return on investment.

[0115] After obtaining the estimated ROI of the target ad in S15, when the target ad is in the data-sparse cold start phase, the multi-dimensional delivery features undergo a second gradient boosting decision processing to obtain an auxiliary ROI. The cold start phase refers to the stage where the target ad has been online for a short time and the cumulative delivery data volume has not reached the preset data volume threshold. In this stage, due to data sparsity, the outputs of network components such as fully connected layers and attention networks, which need to learn parameters from the data, may not be stable enough. The second gradient boosting decision processing involves directly inputting the original multi-dimensional delivery features into an independent gradient boosting decision tree model for prediction. This model does not undergo nonlinear transformations of fully connected layers, attention calculations, or gating calculations; it directly performs gradient boosting decision processing on the original features. The auxiliary ROI refers to the estimated ROI output by the second gradient boosting decision processing, and its role is to provide supplementary reference signals during the cold start phase.

[0116] For example, after obtaining the estimated return on investment (ROI) of the target ad in S15, the launch time and cumulative data volume of the target ad are obtained. If the launch time is less than a preset time threshold and the cumulative data volume does not reach a preset data volume threshold, it is determined that the target ad is in a cold start stage with sparse data. Then, the original multi-dimensional delivery features are directly input into an independent gradient boosting decision tree model for prediction. This model does not go through the nonlinear transformation of the fully connected layer, attention calculation and gating calculation, but directly performs the second gradient boosting decision processing on the original features and outputs the corresponding auxiliary ROI.

[0117] In one embodiment, assuming the target ad has been online for only 2 days, accumulating 5000 impressions, 120 clicks, and 8 activations, and that 2 days is less than a preset time threshold (e.g., 10 days), and the accumulated data volume is far below a preset data volume threshold (e.g., 10000 impressions and 50 activations), then the target ad is determined to be in the cold start phase. Therefore, the original multi-dimensional delivery features [cost=1000, impressions=5000, clicks=120, activations=8, retention rate=0.25, paid=80] are directly input into the second gradient boosting decision tree model, i.e., without FC encoding and attention / gating processing. This second gradient boosting decision tree model outputs an auxiliary ROI based on a simple statistical pattern of the original multi-dimensional delivery features, assuming the output auxiliary ROI is 0.72.

[0118] Through the above implementation method, by introducing an independent gradient boosting decision tree that enhances the path by bypassing the attention-gated neural network in the cold start phase, the defects of unreliable fully connected encoding and attention / gated output caused by data sparsity are overcome. Since the pure gradient boosting decision tree model does not rely on the learning of feature interactions, it can still give a relatively stable prediction based on the statistical pattern of the original features under sparse data. Therefore, the second gradient boosting decision processing provides a safety net for the estimation of the return on investment (ROI) of advertising in the cold start phase, thereby improving the accuracy of the ROI estimation in the cold start phase.

[0119] S32. The estimated return on investment (ROI) and the auxiliary ROI are weighted and combined to obtain the target ROI. The weight of the estimated ROI gradually increases as the cumulative data volume of the target advertisement increases, while the weight of the auxiliary ROI gradually decreases as the cumulative data volume increases.

[0120] After obtaining the auxiliary ROI in S31, the estimated ROI and the auxiliary ROI are weighted and fused to obtain the target ROI. Weighted fusion refers to multiplying each of the two ROI estimates by its respective weight coefficient and then adding them together to obtain the final fused estimate. The weight of the estimated ROI gradually increases as the cumulative data volume of the target advertisement increases, while the weight of the auxiliary ROI gradually decreases as the cumulative data volume increases. The target ROI refers to the final ROI estimate after weighted fusion.

[0121] For example, after obtaining the auxiliary ROI in S31, the weights are first determined based on the current cumulative data volume of the target ad, i.e., the weights of the estimated ROI and the auxiliary ROI. A weight transition function (e.g., linear transition or sigmoid transition) can be preset so that when the cumulative data volume is below the cold start threshold, the weight of the auxiliary ROI is close to 1, and the weight of the estimated ROI is close to 0; when the cumulative data volume reaches the sufficient threshold, the weight of the estimated ROI is close to 1, and the weight of the auxiliary ROI is close to 0; in the intermediate region, the weights transition linearly or smoothly with the data volume. After determining the weights of the estimated ROI and the auxiliary ROI, the estimated ROI is multiplied by its corresponding weight to obtain a first value, and the auxiliary ROI is multiplied by its corresponding weight to obtain a second value. The first and second values ​​are then added together to obtain the target ROI.

[0122] In one embodiment, assume the current cumulative data volume is 5000 impressions, the cold start threshold is 10000 impressions, and the sufficient threshold is 50000 impressions, placing it in a transition range. Assume the auxiliary ROI is determined to be 0.72, and the estimated ROI is 0.88. Following a linear transition, assuming the weight of the auxiliary ROI is determined to be 0.7, and the weight of the estimated ROI is 0.3, the target ROI = 0.88 × 0.3 + 0.72 × 0.7 = 0.768.

[0123] Through the above implementation method, the dynamic weight fusion mechanism that smoothly changes with the accumulation of data enables a smooth transition in the ROI prediction of advertising from the cold start stage to the data-rich stage. This avoids jumps in the prediction value and discontinuity in decision-making during stage switching, thereby improving the overall reliability of the ROI prediction corresponding to advertising.

[0124] S33. Perform placement control operations on target advertisements based on the target return on investment ratio.

[0125] After obtaining the target ROI in S32, a delivery control operation is performed on the target advertisement based on the target ROI. For example, the original estimated ROI is replaced with the target ROI obtained in S32, and the delivery control operation is performed according to the same three-stage decision-making logic as in S16. The specific implementation method for performing the delivery control operation here is the same as that in S16, and will not be repeated here.

[0126] The above-mentioned independent gradient boosting decision (i.e., second gradient boosting decision processing) in the cold start phase, along with the weighted fusion mechanism, solves the problem of unstable fully connected encoding and attention / gating output under sparse data conditions, ensuring the continuity and stability of ROI prediction throughout the entire advertising lifecycle (from zero data cold start to sufficient data).

[0127] Figure 6 This is a flowchart of another advertising delivery method provided in the embodiments of this application, see below. Figure 6 When performing ad delivery control operations, this ad delivery method specifically includes: S166. When the estimated return on investment is lower than the preset optimization trigger threshold, obtain the attention weights corresponding to each feature dimension of the attention network output in the same forward processing that generates the estimated return on investment.

[0128] After obtaining the estimated ROI of the target ad in S15, when the estimated ROI is lower than a preset optimization trigger threshold, the attention weights corresponding to each feature dimension of the attention network output in the same forward processing that generated the estimated ROI are obtained. Here, "same forward processing" refers to the current execution of the complete calculation process from multi-dimensional ad input to estimated ROI output. The optimization trigger threshold is the lowest ROI threshold value that triggers attribution analysis of ad delivery parameters. For example, when the estimated ROI is lower than this threshold, not only is budget reduction executed, but feature attribution analysis is also initiated to find the root cause features leading to the low ROI. These attention weights are calculated in real-time during the generation of the current estimated ROI, reflecting the importance contribution of each feature dimension to the ROI prediction in the current sample.

[0129] For example, after obtaining the estimated return on investment (ROI) of the target ad in the aforementioned S15, assuming that the estimated ROI is 0.68, which is lower than the preset optimization trigger threshold 1, the attention weights of each feature dimension are extracted from the forward processing step S12: [the dimension corresponding to the consumption amount = 0.28, the number of impressions = 0.05, the number of clicks = 0.22, the number of activations = 0.09, the retention rate = 0.25, and the amount paid = 0.11].

[0130] By reusing the attention weights in the same forward processing that generates the estimated return on investment, the importance information of each feature dimension in the current sample can be obtained without additional forward inference. This achieves shared computation for prediction and attribution, reducing the computational overhead of attribution analysis.

[0131] S167. Based on the frequency of use and splitting gain of each feature dimension in the first gradient boosting decision processing when splitting tree nodes, determine the marginal contribution of each feature dimension to the estimated return on investment.

[0132] After obtaining the attention weights in S166, the marginal contribution of each feature dimension to the estimated return on investment is determined based on the frequency of use and splitting gain of each feature dimension during tree node splitting in the first gradient boosting decision processing. Here, frequency of use refers to the total number of times a feature dimension in the gradient boosting decision tree model is selected as a splitting node by various decision trees. Splitting gain refers to the sum of reductions in the loss function (such as mean squared error) when a feature dimension is used as a splitting node. Marginal contribution is an indicator that reflects the degree and direction of influence of the feature dimension on the final estimated return on investment, calculated by comprehensively considering frequency of use and splitting gain. For example, a positive contribution indicates that the value of the feature dimension tends to increase the return on investment, while a negative contribution indicates that the value of the feature dimension tends to decrease the return on investment.

[0133] For example, after obtaining the attention weights in S166, during the gradient boosting decision-making process for the fused feature set, the frequency of use and splitting gain of each feature dimension during tree node splitting are statistically analyzed. Based on the statistically obtained frequency of use and splitting gain of each feature dimension during tree node splitting, the marginal contribution of each feature dimension to the estimated return on investment (ROI) is calculated. For instance, during the gradient boosting decision-making process for the fused feature set, it is found that the dimension corresponding to the consumption amount was used 12 times as a splitting node, with a cumulative splitting gain of 0.45, and the calculated marginal contribution is +0.15, a positive value, indicating a need to increase the ROI; the number of impressions was used 3 times, with a cumulative gain of 0.08, and a marginal contribution of +0.03; the number of clicks was used 15 times, with a cumulative gain of 0.52, and a marginal contribution of -0.22, a negative value, indicating a need to decrease the ROI, meaning that although the number of clicks is high in the current sample, they may be low-quality clicks.

[0134] Through the above implementation method, by using statistical information to determine the marginal contribution of each feature dimension by utilizing the features naturally generated during the splitting process of the gradient boosting decision tree model, no additional model training or feature importance calculation is required, thus achieving information sharing for prediction and attribution.

[0135] S168. The feature dimension with the largest negative marginal contribution is identified as the root cause feature. Based on the covariance of the attention weights of each feature dimension on multiple historical samples of the target advertisement, the associated feature with the highest attention weight covariance with the root cause feature is identified from all feature dimensions other than the root cause feature.

[0136] After determining the marginal contribution of each feature dimension in S167 above, the feature dimension with the largest negative marginal contribution and the largest absolute value is identified as the root cause feature. Based on the covariance of the attention weights of each feature dimension across multiple historical samples of the target advertisement, the associated feature with the highest attention weight covariance with the root cause feature is determined from all feature dimensions other than the root cause feature. Here, the root cause feature refers to the feature dimension that contributes the most negatively to low ROI and is most likely the direct cause of the low ROI. The attention weight covariance refers to the covariance value of the attention weights of two feature dimensions across multiple historical samples. For example, a high covariance indicates that the changes in attention of the two feature dimensions are highly synchronized, meaning there is a strong interaction or correlation between them. The associated feature is the feature with the highest covariance with the root cause feature in the attention dimension, indicating that this feature has the closest synergistic relationship with the root cause feature in the model decision-making process.

[0137] For example, after determining the marginal contribution of each feature dimension in S167, the marginal contribution of each feature dimension is sorted, and the feature dimension with the largest negative absolute value is selected and determined as the root cause feature. For example, the marginal contribution ranking is: click count = -0.22 (largest negative), retention rate = +0.10, spending = +0.15... Therefore, the root cause feature is determined to be click count. After determining the root cause feature, the covariance of the attention weight of each feature dimension on multiple historical samples of the target ad is calculated, and the covariance of the attention weight of each feature dimension other than the root cause feature with respect to the root cause feature is determined. For example, the covariance of the attention weight of click count and the attention weight of the other features on the past 7 days of samples of the target ad is calculated. For example, the covariance of click count and spending is calculated to be 0.45 (highest), the covariance of click count and impression count is 0.12, the covariance of click count and activation count is 0.08... Next, the feature dimension with the highest attention weight covariance relative to the root cause feature among all feature dimensions other than the root cause feature is identified, and this feature dimension is determined as the associated feature. For example, if the calculated associated feature is the spending amount (highest covariance), it means that in the current ad, the attention changes in click count and spending amount are highly correlated, that is, poor click quality leads to low spending efficiency.

[0138] Through the above implementation method, by combining the statistical feature contribution of gradient boosting decision trees and the cross-sample collaborative mode of attention networks, a deep attribution analysis is achieved from "knowing which feature lowers the return on investment" to "knowing which feature is linked to this feature", providing a more comprehensive decision basis for subsequent parameter adjustment.

[0139] S169. Based on root cause characteristics and correlation characteristics, determine the delivery parameters to be adjusted in the budget reduction operation and the adjustment direction of each delivery parameter, and execute the budget reduction operation.

[0140] After determining the root cause and related features in S168, the target delivery parameters to be adjusted and the adjustment direction for each parameter are determined based on these features, and the budget reduction operation is executed. For example, suppose the root cause feature is click count (poor click quality), and the related feature is spend (low spend efficiency). Based on the root cause feature mapping table, the adjustable delivery parameters corresponding to click count are "ad creative" (replace creative to improve click quality) and "targeted audience" (adjust targeting to improve click-through and conversion matching), with adjustment directions of "replace with high-conversion creative" and "tighten audience targeting"; the adjustable delivery parameter corresponding to spend is "bid," with adjustment direction of "lower bid." Then, while sending the budget reduction instruction to the advertising platform, the above-mentioned delivery parameter adjustment suggestions are attached.

[0141] By combining attention weights and the feature contribution information of gradient boosting decision trees through the above implementation methods, a complete diagnostic chain is achieved, from "knowing the return on investment is low" to "knowing why it is low" and then to "knowing what to adjust." Utilizing the reuse of attention weights in the same forward processing and cross-sample attention covariance analysis avoids additional model inference overhead, upgrading budget reduction operations from a broad, one-size-fits-all approach to a refined campaign management system that accurately identifies root causes and adjusts corresponding campaign parameters accordingly.

[0142] The above-mentioned approach combines attention weights and feature contribution information from gradient-enhanced decision trees to achieve a complete intelligent diagnostic chain, from low ROI detection, root cause feature localization, correlation feature discovery to precise adjustment of campaign parameters. This solves the technical problems of existing technologies, such as having only estimated results without attribution analysis, relying on human experience to infer root causes for campaign optimization, and applying budget reduction in a one-size-fits-all manner without targeted parameter adjustments. It upgrades budget reduction operations from extensive budget cuts to precise parameter-level campaign optimization, thereby improving the accuracy and reliability of advertising campaign operations.

[0143] Figure 7 This is a flowchart of another advertising delivery method provided in the embodiments of this application, referred to... Figure 7 When determining the threshold value, this advertising delivery method specifically includes: S131. Determine the current cumulative data volume of each feature dimension in the multi-dimensional delivery feature. For feature dimensions whose cumulative data volume has not reached the preset data volume threshold, mark them as under-observed dimensions.

[0144] When determining the threshold value in S13, the first step is to determine the current cumulative data volume of each feature dimension in the multi-dimensional advertising features. Feature dimensions whose cumulative data volume has not reached the preset data volume threshold are marked as under-observed dimensions. Here, cumulative data volume refers to the number of observations or data samples accumulated for a feature dimension since the target advertisement went live. The preset data volume threshold is the threshold value used to determine whether the data for a feature dimension is sufficient. Under-observed dimensions are feature dimensions whose cumulative data volume is lower than the preset data volume threshold or whose information reliability is questionable due to insufficient observation.

[0145] For example, for each feature dimension in the multi-dimensional delivery feature, query its current cumulative data volume. Compare the cumulative data volume with a preset data volume threshold, and mark those below the threshold as under-observed dimensions.

[0146] In one embodiment, assume the preset data thresholds for expenditure amount are 2000 yuan, impressions are 10000, clicks are 500, activations are 30, retention rate samples are 100 users, and payment amount samples are 50 paying users. A target ad has expenditures of 5000 yuan (met), impressions of 15000 (met), clicks of 350 (not met, marked as an under-observed dimension), activations of 12 (not met, marked as an under-observed dimension), retention rate samples of 80 (not met, marked as an under-observed dimension), and payment amount samples of 30 (not met, marked as an under-observed dimension). The under-observed dimensions for this target ad are clicks, activations, retention rate samples, and payment amount.

[0147] Through the above implementation method, by introducing an under-observed dimension identification mechanism, the sufficiency of data volume is incorporated as an independent judgment dimension into the gating value calculation process, providing a clear labeling basis for subsequent attenuation processing.

[0148] S132. Based on the degree of deviation of the current value of each feature dimension from its value distribution, calculate the initial gate value corresponding to each feature dimension.

[0149] After labeling the underobserved dimensions in S131, the initial gating value for each feature dimension is calculated based on the deviation of its current value from its value distribution. The processing procedure in this step is the same as the calculation method for the deviation and gating value in S13, and will not be repeated here.

[0150] It should be noted that the gating value obtained at this time is called the initial gating value in this step, to distinguish it from the final gating value after the attenuation process in S133 below.

[0151] S133. For a feature dimension marked as an underobserved dimension, multiply its initial gate value by a decay coefficient to obtain the target gate value of the feature dimension. The decay coefficient is positively correlated with the ratio of the cumulative data volume of the feature dimension to the preset data volume threshold and is less than one.

[0152] After obtaining the initial gating value in S132, for the feature dimension marked as underobserved, its initial gating value is multiplied by a decay coefficient to obtain the target gating value of the feature dimension. The decay coefficient is a multiplicative weakening factor applied to the initial gating value of the underobserved dimension. Its value is between 0 and 1 and is positively correlated with the ratio of the cumulative data volume to a preset data volume threshold. For example, the smaller the ratio (the sparser the data), the smaller the decay coefficient (the greater the weakening); the closer the ratio is to 1 (the closer the data is to sufficient), the closer the decay coefficient is to 1 (the less the weakening).

[0153] For example, assuming the cumulative click count is 350 and the preset data threshold is 500, the ratio of the two is 350 / 500 = 0.7. Assuming the decay coefficient function is linear, i.e., decay coefficient = 0.3 + 0.7 × ratio = 0.3 + 0.7 × 0.7 = 0.79. Assuming the initial threshold value for click count is 0.78 (calculated based on deviation in S132), then the target threshold value = 0.78 × 0.79 = 0.62. If the cumulative activation count is 12 and the preset data threshold is 30, the ratio of the two is 0.4, and the decay coefficient = 0.3 + 0.7 × 0.4 = 0.58, assuming the initial threshold value for activation count is 0.90, then the target threshold value = 0.90 × 0.58 = 0.52.

[0154] Through the above implementation method, the initial gate value of the under-observed dimension is further weakened by a decay coefficient based on the amount of accumulated data. This ensures that even if a feature currently has a normal value (i.e., small deviation or high initial gate value), its gate value is still suppressed as long as its data accumulation is insufficient, thus guaranteeing conservatism during the data sparsity phase. As data accumulation increases, the decay gradually dissipates, and the influence of the feature is smoothly restored.

[0155] S134. For feature dimensions that are not labeled as underobserved dimensions, use their initial gate value as the target gate value of the feature dimension.

[0156] After obtaining the initial gate value in S132, for feature dimensions that are not marked as underobserved dimensions, their initial gate value is used as the target gate value for the feature dimension. For example, assuming that the consumption amount and the number of impressions meet the target, and the feature dimension is not marked as underobserved, its target gate value is directly taken as the initial gate value (without additional attenuation).

[0157] In one embodiment, the subsequent S14 performs the corresponding operation with the target gating value of the corresponding dimension.

[0158] As described above, by introducing under-observed dimension identification based on data accumulation and a two-layer gating structure into the gating value calculation—namely, a two-layer gating structure from deviation, initial gating value, under-observed decay to the target gating value—the technical problem of single deviation evaluation being unable to distinguish between two different types of unreliable sources: abnormal values ​​and insufficient data. This embodiment provides a design where the decay coefficient is positively correlated with the data accumulation ratio, ensuring that features with sparser data are decayed more, while features with more abundant data are closer to their natural gating value based on deviation. As the target ad campaign progresses and data accumulates, the decay automatically dissipates, and the feature influence smoothly recovers, achieving a gradual cold start transition within the same attention-gated neural network framework without requiring additional independent models or phased switching logic.

[0159] Based on the above embodiments, Figure 8 This is a schematic diagram of an advertising delivery device provided in an embodiment of this application. (Reference) Figure 8 The advertising delivery device provided in this embodiment specifically includes: a feature acquisition module 21, a dimensionless processing module 22, a latent representation vector calculation module 23, an attention calculation module 24, a gating value determination module 25, a feature enhancement module 26, a feature fusion module 27, and a delivery operation determination module 28.

[0160] Among them, the feature acquisition module 21 is used to acquire the multi-dimensional delivery features of the target advertisement; The dimensionless processing module 22 is used to perform dimensionless processing on the multidimensional projection features to obtain the first multidimensional feature; The latent representation vector calculation module 23 is used to input the first multidimensional feature into at least one fully connected layer and perform a nonlinear transformation to obtain the latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multidimensional feature. Attention calculation module 24 is used to perform attention calculation on the latent representation vector through the attention network to obtain the attention weights corresponding to each feature dimension. The gate value determination module 25 is used to obtain the value distribution of each feature dimension in multiple historical advertisements of the same type as the target advertisement, and determine the gate value corresponding to each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The feature dimension whose current value deviates further from its value distribution has a lower gate value. The feature enhancement module 26 is used to multiply the value of each dimension of the latent representation vector, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension element by element to obtain the enhanced feature vector; The feature fusion module 27 is used to concatenate the multi-dimensional delivery features and the enhanced feature vector along the feature dimension to obtain a fused feature set, and to perform the first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment of the target advertisement. The delivery operation determination module 28 is used to perform delivery control operations on the target advertisement based on the estimated return on investment.

[0161] In one embodiment, the dimensionless processing module 22 includes: a data type identification submodule and a normalization submodule; The data type identification submodule is used to identify the data type of each feature in the multidimensional delivery features. It keeps the original values ​​of numerical features unchanged and maps categorical features to numerical scalars through the embedding layer to obtain a standardized feature set. The normalization submodule is used to normalize the standardized feature set to obtain the first multidimensional feature.

[0162] Based on the above implementation, the feature fusion module 27 is also used to sequentially perform first gradient boosting decision processing on multiple preset time nodes in order from the nearest to the farthest time node, so as to obtain the estimated production ratio corresponding to each preset time node; wherein, the processing of the first preset time node takes the fused feature set as input, and the processing of each subsequent preset time node takes the feature set obtained by splicing the fused feature set with the estimated production ratio of the previously processed preset time node as input, so that the estimated production ratio of the subsequent preset time node is based on the estimated result of the previous preset time node; In one embodiment, the deployment operation determination module 28 includes: a trend determination submodule, a deployment operation determination submodule, and a deployment execution submodule; The trend determination submodule is used to calculate the trend direction confidence based on the consistency of the change direction of the estimated output ratio between adjacent preset time nodes. The trend direction confidence is the proportion of the number of intervals where adjacent preset time nodes change in the same direction to the total number of intervals. The deployment operation determination submodule is used to determine the deployment control operation based on the trend direction of the time-series trajectory formed by the trend direction confidence level and the estimated return ratio at each preset time node; The deployment execution submodule is used to execute the following operations: when the time-series trajectory shows an upward trend and the estimated return on investment (ROI) at the most recent preset time node is greater than or equal to a first preset threshold and the confidence level of the trend direction is greater than or equal to a preset direction confidence threshold, it executes the continued deployment operation; when the time-series trajectory shows an upward trend but the estimated ROI at the most recent preset time node is less than the first preset threshold, it executes the budget reduction operation; when the time-series trajectory shows an upward trend and the estimated ROI at the most recent preset time node is greater than or equal to the first preset threshold, but the confidence level of the trend direction is less than the preset direction confidence threshold, it executes the budget reduction operation and extends the data observation period; when the time-series trajectory shows a downward trend or a stable trend and the estimated ROI at the most recent preset time node is less than a second preset threshold, it executes the shutdown operation; when the time-series trajectory shows a downward trend or a stable trend and the estimated ROI at the most recent preset time node is greater than or equal to the second preset threshold, it executes the budget reduction operation.

[0163] In one embodiment, the delivery operation determination module 28 includes: a delivery stage determination submodule, a stage threshold determination submodule, a first execution submodule, a second execution submodule, and a third execution submodule; The campaign stage determination submodule is used to determine the current campaign stage of the target ad. The current campaign stage is determined based on the proportion of the target ad's cumulative spending budget to the preset total budget. When the spending ratio is lower than the first spending ratio threshold, the current campaign stage is determined to be the first stage. When the spending ratio is between the first spending ratio threshold and the second spending ratio threshold, the current campaign stage is determined to be the second stage. When the spending ratio is higher than the second spending ratio threshold, the current campaign stage is determined to be the third stage. The stage threshold determination submodule is used to determine the corresponding upper and lower thresholds of the return on investment (ROI) based on the current deployment stage. Specifically, the upper and lower thresholds of the ROI in the first stage are lower than the corresponding thresholds in the second stage, and the upper and lower thresholds of the ROI in the third stage are higher than the corresponding thresholds in the second stage. The first execution submodule is used to continue the delivery of the target ad when the estimated return on investment is greater than or equal to the upper limit threshold of the return on investment corresponding to the current delivery stage. The second execution submodule is used to perform a budget reduction operation on the target ad when the estimated return on investment is less than the upper limit threshold of the return on investment corresponding to the current campaign stage and greater than or equal to the lower limit threshold of the return on investment corresponding to the current campaign stage. The third execution submodule is used to shut down the target ad when the estimated return on investment is less than the lower limit threshold of the return on investment for the current campaign stage.

[0164] In one embodiment, the advertising delivery device further includes: an auxiliary processing module and a weighted fusion module; The auxiliary processing module is used to perform a second-gradient improvement decision processing on the multi-dimensional delivery features when the target advertisement is in the cold start stage with sparse data, so as to obtain an auxiliary return on investment. The weighted fusion module is used to weight and fuse the estimated return on investment (ROI) and the auxiliary ROI to obtain the target ROI. The weight of the estimated ROI gradually increases as the cumulative data volume of the target advertisement increases, while the weight of the auxiliary ROI gradually decreases as the cumulative data volume increases. Based on the aforementioned implementation, the placement operation determination module 28 is also used to perform placement control operations on the target advertisement according to the target return on investment ratio.

[0165] In one embodiment, the delivery operation determination module 28 includes: a forward acquisition submodule, a marginal contribution determination submodule, a first feature determination submodule, and a fourth execution submodule; The forward acquisition submodule is used to acquire the attention weights corresponding to each feature dimension of the attention network output in the same forward processing that generates the estimated return on investment when the estimated return on investment is lower than the preset optimization trigger threshold. The marginal contribution determination submodule is used to determine the marginal contribution of each feature dimension to the estimated return ratio based on the frequency of use and splitting gain of each feature dimension in the first gradient improvement decision processing when splitting tree nodes. The first feature determination submodule is used to determine the feature dimension with the largest negative marginal contribution as the root cause feature. Based on the covariance of the attention weight of each feature dimension on multiple historical samples of the target advertisement, it determines the associated feature with the highest attention weight covariance with the root cause feature from all feature dimensions other than the root cause feature. The fourth execution submodule is used to determine the delivery parameters to be adjusted in the budget reduction operation and the adjustment direction of each delivery parameter based on the root cause characteristics and correlation characteristics, and then execute the budget reduction operation.

[0166] In one embodiment, the gate value determination module 25 includes: a dimension marking submodule, an initial gate calculation submodule, a first gate determination submodule, and a second gate determination submodule; The dimension labeling submodule is used to determine the current cumulative data volume of each feature dimension in the multi-dimensional delivery feature. For feature dimensions whose cumulative data volume has not reached the preset data volume threshold, they are labeled as under-observed dimensions. The initial gating calculation submodule is used to calculate the initial gating value for each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The first gating determination submodule is used to multiply the initial gating value of the feature dimension marked as underobserved by a decay coefficient to obtain the target gating value of the feature dimension. The decay coefficient is positively correlated with the ratio of the cumulative data volume of the feature dimension to the preset data volume threshold and is less than one. The second gating determination submodule is used to use the initial gating value of the feature dimension that is not marked as an underobserved dimension as the target gating value of the feature dimension.

[0167] The advertising delivery device provided in this application embodiment can be used to execute the advertising delivery method provided in the above embodiment, and has corresponding functions and beneficial effects.

[0168] This application provides an advertising delivery device, referring to... Figure 9 The advertising delivery device includes a processor 31, a memory 32, a communication module 33, an input device 34, and an output device 35. The advertising delivery device may have one or more processors, and one or more memories. The processor, memory, communication module, input device, and output device of the advertising delivery device can be connected via a bus or other means.

[0169] The memory 32, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as program instructions / modules corresponding to the advertising delivery method in any embodiment of this application (e.g., feature acquisition module, dimensionless processing module, implicit representation vector calculation module, attention calculation module, gating value determination module, feature enhancement module, feature fusion module, and delivery operation determination module in the advertising delivery device). The memory may mainly include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store data created according to the use of the device, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the device via a network. Examples of the above-mentioned networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0170] The communication module 33 is used for data transmission.

[0171] The processor 31 executes various functional applications and data processing of the device by running software programs, instructions and modules stored in the memory, thereby realizing the above-mentioned advertising delivery method.

[0172] Input device 34 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the device. Output device 35 may include display devices such as a display screen.

[0173] The advertising delivery equipment provided above can be used to execute the advertising delivery method provided in the above embodiments, and has corresponding functions and beneficial effects.

[0174] This application embodiment also provides a storage medium for storing computer-executable instructions. When executed by a computer processor, the computer-executable instructions are used to execute an advertising delivery method. The advertising delivery method includes: acquiring multi-dimensional delivery features of a target advertisement; performing dimensionless processing on the multi-dimensional delivery features to obtain a first multi-dimensional feature; inputting the first multi-dimensional feature into at least one fully connected layer for nonlinear transformation to obtain a latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multi-dimensional feature; performing attention calculation on the latent representation vector through an attention network to obtain the attention weights corresponding to each feature dimension; and acquiring the attention weights of each feature dimension in the same type as the target advertisement. The value distribution of multiple historical advertisements is analyzed. Based on the deviation of the current value of each feature dimension from its value distribution, the gate value corresponding to each feature dimension is determined. The feature dimension whose current value deviates further from the value distribution has a lower gate value. The value of each dimension in the latent representation vector, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension are multiplied element-wise to obtain the enhanced feature vector. The multi-dimensional delivery features and the enhanced feature vector are concatenated along the feature dimensions to obtain the fused feature set. The fused feature set is then subjected to the first gradient boosting decision processing to obtain the estimated return on investment (ROI) of the target advertisement. The delivery control operation of the target advertisement is performed based on the estimated ROI.

[0175] Storage medium – any type of memory device or storage device. The term “storage medium” is intended to include: mounting media, such as CD-ROM, floppy disk, or magnetic tape devices; computer system memory or random access memory, such as DRAM, DDR RAM, SRAM, EDO RAM, Rambus RAM, etc.; non-volatile memory, such as flash memory, magnetic media (e.g., hard disk or optical storage); registers or other similar types of memory elements, etc. Storage medium may also include other types of memory or combinations thereof. Furthermore, storage medium may reside in a first computer system in which the program is executed, or it may reside in a different second computer system connected to the first computer system via a network (such as the Internet). The second computer system can provide program instructions to the first computer for execution. The term “storage medium” can include two or more storage media residing in different locations (e.g., in different computer systems connected via a network). Storage medium may store program instructions (e.g., specifically implemented as a computer program) executable by one or more processors.

[0176] Of course, the computer-executable instructions stored in the storage medium provided in the embodiments of this application are not limited to the advertising delivery method described above, but can also perform related operations in the advertising delivery method provided in any embodiment of this application.

[0177] The advertising delivery device, storage medium, and advertising delivery equipment provided in the above embodiments can execute the advertising delivery method provided in any embodiment of this application. For technical details not described in detail in the above embodiments, please refer to the advertising delivery method provided in any embodiment of this application.

[0178] The above description is merely a preferred embodiment and the technical principles employed in this application. This application is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions that can be made by those skilled in the art will not depart from the scope of protection of this application. Therefore, although this application has been described in detail through the above embodiments, this application is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of this application, the scope of which is determined by the scope of the claims.

Claims

1. An advertising placement method, characterized in that, include: The multidimensional delivery features of the target advertisement are obtained, and the multidimensional delivery features are dedimensionalized to obtain the first multidimensional feature. The first multidimensional feature is input into at least one fully connected layer for nonlinear transformation to obtain the latent representation vector. The dimension of the latent representation vector is the same as the dimension of the first multidimensional feature. Attention weights for each feature dimension are obtained by performing attention calculations on the latent representation vector using an attention network. Obtain the value distribution of each feature dimension in multiple historical advertisements of the same type as the target advertisement, and determine the gate value corresponding to each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The feature dimension whose current value deviates further from the value distribution has a lower gate value. The value of each dimension of the latent representation vector, the attention weight of the corresponding dimension, and the gating value of the corresponding dimension are multiplied element-wise to obtain the enhanced feature vector. The multi-dimensional delivery features and the enhanced feature vector are concatenated along the feature dimension to obtain a fused feature set. The fused feature set is then subjected to first gradient boosting decision processing to obtain the estimated return on investment of the target advertisement. The delivery control operation is performed on the target advertisement based on the estimated return on investment.

2. The method according to claim 1, characterized in that, The step of performing dimensionless processing on the multidimensional projection features to obtain the first multidimensional feature includes: Identify the data type of each feature in the multidimensional delivery features, keep the original values ​​of numerical features unchanged, and map categorical features to numerical scalars through an embedding layer to obtain a standardized feature set; The standardized feature set is normalized to obtain the first multidimensional feature.

3. The method according to claim 1, characterized in that, The step of performing a first gradient boosting decision process on the fused feature set to obtain the estimated return on investment for the target advertisement includes: In order of time nodes from nearest to farthest, multiple preset time nodes are sequentially subjected to first gradient boosting decision processing to obtain the estimated output ratio corresponding to each preset time node; wherein, the processing of the first preset time node is based on the fused feature set as input, and the processing of each subsequent preset time node is based on the feature set obtained by concatenating the fused feature set with the estimated output ratio of the previously processed preset time node as input, so that the estimated output ratio of the subsequent preset time node is based on the estimated results of the previous preset time node; Accordingly, the step of performing delivery control operations on the target advertisement based on the estimated return on investment includes: The confidence level of the trend direction is calculated based on the consistency of the change direction of the estimated production ratio between adjacent preset time nodes. The confidence level of the trend direction is the proportion of the number of intervals where adjacent preset time nodes change in the same direction to the total number of intervals. The deployment control operation is determined based on the trend direction of the time-series trajectory formed by the confidence level of the trend direction and the estimated input-output ratio at each preset time node. When the time-series trajectory shows an upward trend and the estimated return on investment at the most recent preset time node is greater than or equal to a first preset threshold, and the confidence level of the trend direction is greater than or equal to a preset direction confidence threshold, continue the deployment operation; when the time-series trajectory shows an upward trend but the estimated return on investment at the most recent preset time node is less than the first preset threshold, reduce the budget operation; when the time-series trajectory shows an upward trend and the estimated return on investment at the most recent preset time node is greater than or equal to the first preset threshold, but the confidence level of the trend direction is less than the preset direction confidence threshold, reduce the budget operation and extend the data observation period; when the time-series trajectory shows a downward trend or a stable trend and the estimated return on investment at the most recent preset time node is less than a second preset threshold, shut down the operation; when the time-series trajectory shows a downward trend or a stable trend and the estimated return on investment at the most recent preset time node is greater than or equal to the second preset threshold, reduce the budget operation.

4. The method according to claim 1, characterized in that, The step of performing delivery control operations on the target advertisement based on the estimated return on investment includes: The current campaign stage of the target advertisement is determined based on the proportion of the cumulative spending budget of the target advertisement to the preset total budget. When the spending proportion is lower than a first spending proportion threshold, the current campaign stage is determined to be a first stage. When the spending proportion is between the first spending proportion threshold and a second spending proportion threshold, the current campaign stage is determined to be a second stage. When the spending proportion is higher than the second spending proportion threshold, the current campaign stage is determined to be a third stage. The upper and lower limits of the return on investment ratio are determined according to the current deployment stage. The upper and lower limits of the return on investment ratio in the first stage are lower than the corresponding thresholds in the second stage, and the upper and lower limits of the return on investment ratio in the third stage are higher than the corresponding thresholds in the second stage. When the estimated return on investment is greater than or equal to the upper limit threshold of the return on investment corresponding to the current stage of the campaign, the campaign will continue to be launched for the target ad. When the estimated return on investment (ROI) is less than the upper limit threshold of ROI corresponding to the current campaign stage and greater than or equal to the lower limit threshold of ROI corresponding to the current campaign stage, a budget reduction operation is performed on the target advertisement. When the estimated return on investment is less than the lower limit threshold of the return on investment for the current campaign stage, the target advertisement is shut down.

5. The method according to claim 1, characterized in that, After obtaining the estimated return on investment for the target advertisement, the process further includes: When the target advertisement is in the cold start stage with sparse data, the multi-dimensional delivery features are subjected to a second gradient improvement decision processing to obtain an auxiliary return on investment. The estimated return on investment (ROI) and the auxiliary ROI are weighted and fused to obtain the target ROI. The weight of the estimated ROI gradually increases as the cumulative data volume of the target advertisement increases, while the weight of the auxiliary ROI gradually decreases as the cumulative data volume increases. Accordingly, the step of performing delivery control operations on the target advertisement based on the estimated return on investment includes: The target advertisement is subject to delivery control operations based on the target return on investment.

6. The method according to claim 1, characterized in that, The step of performing delivery control operations on the target advertisement based on the estimated return on investment includes: When the estimated return on investment is lower than the preset optimization trigger threshold, the attention weights corresponding to each feature dimension of the attention network output in the same forward processing that generates the estimated return on investment are obtained. Based on the frequency of use and splitting gain of each feature dimension in the first gradient boosting decision processing when splitting tree nodes, the marginal contribution of each feature dimension to the estimated return on investment is determined. The feature dimension with the largest negative marginal contribution is identified as the root cause feature. Based on the covariance of the attention weights of each feature dimension over multiple historical samples of the target advertisement, the associated feature with the highest attention weight covariance with the root cause feature is determined from the feature dimensions other than the root cause feature. Based on the root cause features and the correlation features, determine the delivery parameters to be adjusted in the budget reduction operation and the adjustment direction of each delivery parameter, and then execute the budget reduction operation.

7. The method according to claim 1, characterized in that, The process of determining the gate value for each feature dimension based on the deviation of the current value of each feature dimension from its value distribution includes: Determine the current cumulative data volume of each feature dimension in the multidimensional delivery feature; for feature dimensions whose cumulative data volume has not reached the preset data volume threshold, mark them as under-observed dimensions. Based on the degree of deviation of the current value of each feature dimension from its value distribution, calculate the initial gate value corresponding to each feature dimension; For a feature dimension marked as an underobserved dimension, its initial gate value is multiplied by a decay coefficient to obtain the target gate value of the feature dimension. The decay coefficient is positively correlated with the ratio of the cumulative data volume of the feature dimension to the preset data volume threshold and is less than one. For feature dimensions that are not labeled as underobserved dimensions, their initial gating value is used as the target gating value for the feature dimension.

8. An advertising delivery device, characterized in that, include: The feature acquisition module is used to acquire multi-dimensional delivery features of the target advertisement; The dimensionless processing module is used to perform dimensionless processing on the multidimensional projection features to obtain the first multidimensional features; The latent representation vector calculation module is used to input the first multidimensional feature into at least one fully connected layer and perform a nonlinear transformation to obtain a latent representation vector, wherein the dimension of the latent representation vector is the same as the dimension of the first multidimensional feature. The attention calculation module is used to perform attention calculation on the latent representation vector through the attention network to obtain the attention weights corresponding to each feature dimension. The gate value determination module is used to obtain the value distribution of each feature dimension in multiple historical advertisements of the same type as the target advertisement, and determine the gate value corresponding to each feature dimension based on the degree of deviation of the current value of each feature dimension from its value distribution. The feature dimension whose current value deviates further from the value distribution has a lower gate value. The feature enhancement module is used to multiply the value of each dimension, the attention weight of the corresponding dimension, and the gate value of the corresponding dimension in the latent representation vector element by element to obtain the enhanced feature vector. The feature fusion module is used to concatenate the multi-dimensional delivery features and the enhanced feature vector along the feature dimension to obtain a fused feature set, and to perform a first gradient boosting decision processing on the fused feature set to obtain the estimated return on investment of the target advertisement; The delivery operation determination module is used to perform delivery control operations on the target advertisement based on the estimated return on investment.

9. An advertising delivery device, characterized in that, include: Memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A storage medium for storing computer-executable instructions, characterized in that, The computer-executable instructions, when executed by a processor, are used to perform the method as described in any one of claims 1-7.