Intelligent bidding method and apparatus based on apple ad marketing interface

Abstract

The application discloses an intelligent bidding method and device based on an apple advertising marketing interface, relates to the technical field of advertisement putting, and comprises the following steps: a double-branch machine learning model containing a long short-term memory time sequence branch and a gradient boosting tree feature branch is constructed; the model is trained by using historical putting data to obtain a target prediction model; real-time putting data is collected through the apple advertising marketing interface, and a plurality of candidate bidding strategies are generated according to preset business rules; the real-time data and the candidate strategies are input into the target prediction model to obtain a conversion probability prediction value and a cost distribution prediction value of the candidate bidding strategies; based on the prediction values, a score function is used to calculate a strategy score value, and a high-score strategy combination is screened out; keywords, audiences, time periods and regional adjustment parameters of the strategy combination are integrated to generate an optimal bidding strategy package. The application can efficiently and automatically optimize the Apple Ads advertising bidding strategy to improve putting efficiency and cost control.

Description

Technical Field

[0001] This application relates to the field of advertising delivery technology, and in particular to a smart bidding method and apparatus based on Apple's advertising marketing interface. Background Technology

[0002] In the internet advertising sector, especially in Apple Ads delivery services for mobile phone manufacturers, ad bidding strategy formulation is a core element affecting ad performance. With increasing market competition and more complex advertising needs, advertisers have higher requirements for efficiency, cost control, and return on investment (ROI). To meet the demands of large-scale, multi-dimensional ad delivery while coping with the dynamic changes in the real-time bidding environment, advertisers urgently need an intelligent system capable of efficiently and accurately optimizing bidding strategies to improve the overall effectiveness of ad delivery.

[0003] Currently, bidding operations based on Apple Ads in the industry mainly rely on optimization specialists to manually analyze data (such as clicks, conversion rates, cost distribution, etc.), manually adjust bidding parameters (such as keyword bids, audience bids, budget allocation, etc.), and set up campaign strategies based on experience.

[0004] However, this manual bidding method has many shortcomings. First, manually analyzing massive amounts of campaign data and formulating bidding strategies is time-consuming and labor-intensive. For example, it can take an optimizer several hours to adjust the bidding parameters of 10 ad groups, which is extremely inefficient. Second, bidding strategies rely on the optimizer's subjective experience and lack data-driven scientific rigor, making it difficult to accurately match Apple Ads' real-time bidding mechanism. This often results in overbidding, leading to cost overruns, or underbidding, causing missed traffic. Furthermore, manual operation cannot respond in real-time to dynamic factors such as traffic fluctuations and competitor bidding changes on the Apple Ads platform. For example, if traffic costs decrease at night, untimely manual adjustments can lead to wasted budget. Finally, when multiple accounts and ad groups are running in parallel, manual management costs increase exponentially, making it difficult to meet the large-scale, multi-dimensional campaign needs of mobile phone manufacturers. Therefore, how to efficiently and automatically optimize Apple Ads ad bidding strategies to improve campaign efficiency and cost control has become an urgent problem to be solved.

[0005] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The purpose of this application is to provide a smart bidding method and device based on the Apple Ads marketing interface, which aims to solve the technical problem of how to efficiently and automatically optimize Apple Ads advertising bidding strategies to improve campaign efficiency and cost control.

[0007] To achieve the above objectives, this application proposes a smart bidding method based on Apple's advertising marketing interface, the method comprising:

[0008] Construct a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch;

[0009] The dual-branch machine learning model is trained using historical delivery data to obtain a target prediction model;

[0010] Real-time campaign data is collected through Apple Ads marketing API, and multiple candidate bidding strategies are generated based on preset business rules;

[0011] The real-time delivery data and the candidate bidding strategies are processed by the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies.

[0012] Based on the predicted conversion probability and the predicted cost distribution, the strategy score of the candidate bidding strategy is calculated using a scoring function, and the strategy combination is determined based on the strategy score.

[0013] By integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the aforementioned strategy combination, an optimal bidding strategy package is generated.

[0014] In one embodiment, the step of constructing a two-branch machine learning model comprising a long short-term memory (LSM) temporal branch and a gradient boosting tree feature branch includes: establishing a LSM temporal branch, which processes time-series data including keyword click-through rate sequences, conversion rate sequences, and competitor bidding sequences, and updates parameters using an adaptive moment estimation optimizer; establishing a gradient boosting tree feature branch, which processes static feature data including keyword industry classification features, audience geographic tag features, audience age tag features, and ad group budget features, and constructs a decision tree using a Friedman gradient descent optimizer; establishing a fully connected fusion layer to connect the output nodes of the LSM temporal branch and the gradient boosting tree feature branch; constructing a weighted mixed loss function based on a binary cross-entropy loss function, a mean squared error loss function, and preset weight coefficients; and integrating the LSM temporal branch, the gradient boosting tree feature branch, the fully connected fusion layer, and the weighted mixed loss function to obtain the two-branch machine learning model.

[0015] In one embodiment, the target prediction model includes a long short-term memory temporal branch, a gradient boosting tree feature branch, and a fully connected fusion layer. The real-time delivery data includes temporal features and static features. The step of processing the real-time delivery data and the candidate bidding strategy through the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy includes: performing time series encoding on the temporal features through the long short-term memory temporal branch to obtain a cost trend vector; performing feature combination correlation analysis on the static features and the adjustment parameters in the candidate bidding strategy through the gradient boosting tree feature branch to obtain a conversion correlation matrix; performing feature fusion on the cost trend vector and the conversion correlation matrix through the fully connected fusion layer to obtain a joint prediction feature; and performing probability distribution transformation and cost quantile mapping on the joint prediction feature to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy.

[0016] In one embodiment, the static features include keyword industry classification features, audience geographic tag features, and audience age tag features; the adjustment parameters in the candidate bidding strategy include keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters; the step of performing feature combination association analysis on the static features and the adjustment parameters in the candidate bidding strategy through the gradient boosting tree feature branches to obtain a conversion association matrix includes: performing feature encoding on the keyword industry classification features, the audience geographic tag features, and the audience age tag features to obtain a one-hot encoded vector; and performing feature encoding on the keyword adjustment parameters, the audience adjustment parameters, and the... The time-period adjustment parameters and the region adjustment parameters are normalized to obtain normalized adjustment parameters; the one-hot encoded vector and the normalized adjustment parameters are concatenated to form an enhanced feature vector; the enhanced feature vector is split and decided through a multi-layer decision tree of the gradient boosting tree feature branches to obtain a feature combination rule set; the weight values ​​of the leaf nodes of the multi-layer decision tree are extracted as combination influence factors; a transformation correlation matrix is ​​constructed based on the feature combination rule set and the combination influence factors, where the rows of the transformation correlation matrix represent feature combination categories, the columns of the transformation correlation matrix represent adjustment parameter categories, and the element values ​​of the transformation correlation matrix represent the change in transformation probability.

[0017] In one embodiment, the step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package includes: performing dimensional separation processing on the strategy combination to obtain a keyword adjustment set, an audience adjustment set, a time period adjustment set, and a geographic adjustment set; performing weighted aggregation on the keyword adjustment set, calculating a weighted average based on the strategy score values ​​of each strategy in the strategy combination to obtain the target keyword adjustment parameters; selecting the top M audience adjustment parameters with the highest strategy score values ​​from the audience adjustment set; performing trend analysis on the time period adjustment set and the geographic adjustment set, generating a dynamic coefficient curve based on a time series pattern; constructing a structured strategy package based on the target keyword adjustment parameters, the top M audience adjustment parameters, and the dynamic coefficient curve; and encapsulating the structured strategy package using the Apple Ads API protocol to obtain the optimal bidding strategy package.

[0018] In one embodiment, the step of generating multiple candidate bidding strategies according to preset business rules includes: parsing the preset business rules to obtain a set of step size parameters for keyword adjustment dimension, audience adjustment dimension, time period adjustment dimension, and region adjustment dimension; generating a discrete adjustment value matrix based on the step size parameter set, wherein the keyword adjustment dimension and the audience adjustment dimension generate percentage adjustment values, and the time period adjustment dimension and the region adjustment dimension generate coefficient adjustment values; combining the adjustment values ​​of different dimensions in the discrete adjustment value matrix by Cartesian product to obtain an initial strategy combination set; and filtering the initial strategy combination set according to preset business constraints to obtain multiple candidate bidding strategies.

[0019] In one embodiment, the historical delivery data includes input data and output labels. The input data includes time-series data and static feature data, and the output labels include conversion rate and cost value. The step of training the dual-branch machine learning model using the historical delivery data to obtain the target prediction model includes: dividing the time-series data into a sliding window according to a preset window length and a preset step size to obtain time-series training samples; performing one-hot encoding and normalization on the static feature data to obtain static feature training samples; inputting the time-series training samples and the static feature training samples into the dual-branch machine learning model to obtain prediction data; calculating a hybrid loss function value based on the prediction data, the conversion rate, and the cost value; and training the dual-branch machine learning model using backpropagation based on the hybrid loss function value to obtain the target prediction model.

[0020] In one embodiment, after the step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package, the method further includes: executing the optimal bidding strategy package through the Apple Ads marketing interface and monitoring the actual conversion cost after executing the optimal bidding strategy package; triggering a bidding strategy rollback operation when the actual conversion cost exceeds a preset cost threshold; and sending an abnormal strategy report to a preset ad placement console device when the bidding strategy rollback operation is triggered a preset number of times consecutively.

[0021] In one embodiment, after the step of executing the optimal bidding strategy package through the Apple Ads marketing interface, the method further includes: monitoring the actual conversion probability after executing the optimal bidding strategy package; when the deviation between the actual conversion probability and the predicted conversion probability value is greater than a preset deviation threshold, returning to the step of training the dual-branch machine learning model using historical delivery data to obtain a target prediction model; when the deviation between the actual conversion probability and the predicted conversion probability value is less than or equal to the preset deviation threshold, generating an ad optimization report based on the optimal bidding strategy package, the actual conversion cost, and the actual conversion probability, including keyword performance comparison dimensions, audience performance comparison dimensions, time period performance comparison dimensions, and geographic performance comparison dimensions.

[0022] Furthermore, to achieve the above objectives, this application also proposes a smart bidding device based on the Apple Ads marketing interface, the device comprising:

[0023] The model building module is used to build a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch.

[0024] The model training module is used to train the dual-branch machine learning model using historical deployment data to obtain the target prediction model;

[0025] The data acquisition module is used to collect real-time campaign data through the Apple Ads marketing interface and generate multiple candidate bidding strategies based on preset business rules.

[0026] The prediction module is used to process the real-time delivery data and the candidate bidding strategies through the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies.

[0027] The strategy scoring module is used to calculate the strategy score of the candidate bidding strategy based on the conversion probability prediction value and the cost distribution prediction value through a scoring function, and to determine the strategy combination based on the strategy score value;

[0028] The strategy determination module is used to integrate the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters of the strategy combination to generate the optimal bidding strategy package.

[0029] In addition, to achieve the above objectives, this application also proposes a smart bidding device based on the Apple Ads Marketing Interface, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the smart bidding method based on the Apple Ads Marketing Interface as described above.

[0030] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the smart bidding method based on the Apple advertising marketing interface as described above.

[0031] One or more technical solutions proposed in this application have at least the following technical effects:

[0032] First, a two-branch machine learning model is constructed, incorporating both a long short-term memory temporal branch and a gradient boosting tree feature branch. This structure fully leverages the advantages of both, improving the model's predictive accuracy and generalization ability. Next, the two-branch machine learning model is trained using historical campaign data. By optimizing the loss function and adjusting the model parameters, a target prediction model capable of accurately predicting conversion probabilities and cost distributions is obtained. Then, real-time campaign data is collected through the Apple Ads marketing interface, and multiple candidate bidding strategies are generated according to preset business rules. Real-time data reflects current market dynamics, providing multiple options for subsequent optimization. Afterward, the real-time campaign data and candidate bidding strategies are input into the target prediction model. The model outputs the conversion probability prediction and cost distribution prediction values ​​for each candidate bidding strategy, allowing for early evaluation of the effectiveness and cost of different bidding strategies and helping the system select the optimal bidding strategy. Subsequently, based on the conversion probability prediction and cost distribution prediction values, a scoring function is used to calculate the strategy score for each candidate bidding strategy. The strategy combination is then determined based on the score. The scoring function balances conversion probability and cost distribution, ensuring that the selected strategy can both improve conversion rates and control costs. Finally, by integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters in the strategy combination, an optimal bidding strategy package is generated. This strategy package can comprehensively guide the advertising system to dynamically adjust bids in a real-time bidding environment, improve the accuracy and efficiency of advertising, and thus efficiently and automatically optimize Apple Ads advertising bidding strategies to improve campaign efficiency and cost control. Attached Figure Description

[0033] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating an embodiment of the smart bidding method based on the Apple Ads marketing interface provided in this application.

[0036] Figure 2 This is a schematic diagram of the strategy rollback process provided in Embodiment 1 of the smart bidding method based on the Apple Ads Marketing Interface of this application;

[0037] Figure 3 This is a flowchart illustrating Embodiment 2 of the smart bidding method based on the Apple Ads Marketing Interface in this application.

[0038] Figure 4 This is a schematic diagram of the core module interaction provided in Embodiment 2 of the smart bidding method based on the Apple Ads marketing interface of this application;

[0039] Figure 5 A simplified flowchart illustrating the smart bidding method based on the Apple Ads marketing interface provided in Embodiment 2 of this application;

[0040] Figure 6 This is a schematic diagram of the module structure of the smart bidding device based on the Apple Ads Marketing Interface in an embodiment of this application.

[0041] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0042] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0043] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0044] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device or intelligent bidding system capable of realizing the above functions. The following description uses an intelligent bidding system as an example to illustrate this embodiment and the subsequent embodiments.

[0045] Based on this, the embodiments of this application provide a smart bidding method based on Apple's advertising marketing interface, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the smart bidding method based on Apple's advertising marketing interface.

[0046] In this embodiment, the smart bidding method based on the Apple advertising marketing interface includes steps S10 to S60:

[0047] Step S10: Construct a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch.

[0048] It should be noted that the Long Short-Term Memory (LSTM) temporal branch is a model structure based on the LSTM network, specifically designed for processing and analyzing data with time-series characteristics. It can effectively capture long-term dependencies and temporal dynamics in the data.

[0049] Gradient Boosting Tree Feature Branch is a model structure based on Gradient Boosting Tree (GBT). It is mainly used to process and analyze data with static features, and improves the accuracy and stability of prediction by combining multiple decision tree models.

[0050] A dual-branch machine learning model is a composite model that combines long short-term memory temporal branches and gradient boosting tree feature branches, aiming to leverage the strengths of different models to process different types of data and features.

[0051] As an example, the steps of constructing a two-branch machine learning model comprising a long short-term memory (LSM) temporal branch and a gradient boosting tree feature branch include: establishing a LSM temporal branch, which processes time-series data including keyword click-through rate sequences, conversion rate sequences, and competitor bidding sequences, and updates parameters using an adaptive moment estimation optimizer; establishing a gradient boosting tree feature branch, which processes static feature data including keyword industry classification features, audience geographic tag features, audience age tag features, and ad group budget features, and constructs a decision tree using a Friedman gradient descent optimizer; establishing a fully connected fusion layer to connect the output nodes of the LSM temporal branch and the gradient boosting tree feature branch; constructing a weighted mixed loss function based on a binary cross-entropy loss function, a mean squared error loss function, and preset weight coefficients; and integrating the LSM temporal branch, the gradient boosting tree feature branch, the fully connected fusion layer, and the weighted mixed loss function to obtain the two-branch machine learning model.

[0052] Keyword click volume sequence refers to the number of times a keyword is clicked by users within a specific time period (such as hourly or daily); conversion rate sequence refers to the percentage of users who click on an ad and complete the expected behavior (such as purchase or registration) within the same time period; competitor bidding sequence refers to the bidding records of competitors for similar keywords within a corresponding time period.

[0053] The Adaptive Moment Estimation (Adam) optimizer is an optimization algorithm for training deep learning models. It combines the advantages of momentum optimization and adaptive learning rate adjustment, dynamically adjusting the learning rate based on historical gradient information of the parameters, thereby improving the convergence speed and stability of model training. The Friedman Gradient Descent optimizer is an optimization algorithm for training GBT models. It is based on the idea of ​​gradient boosting and minimizes the loss function by progressively building a decision tree.

[0054] Keyword industry classification features refer to the industry category to which the keyword belongs (such as "smartphones," "digital products," etc.), used to reflect the business relevance of the keyword; audience geographic tag features refer to the geographical location of the target audience (such as "first-tier cities," "second-tier cities," etc.), used to analyze the impact of regional differences on advertising effectiveness; audience age tag features refer to the age range of the target audience (such as "18-24 years old," "25-35 years old," etc.), used to analyze the response differences of users in different age groups; ad group budget features refer to the budget amount allocated to a specific ad group, used to evaluate the constraints and impact of the budget on advertising effectiveness.

[0055] A fully connected fusion layer is a neural network structure used to integrate the outputs of different branch models. The Binary Cross-Entropy Loss sub-function measures the model's prediction accuracy for binary classification problems (such as whether to convert), evaluating its classification performance by calculating the difference between the predicted probability and the true label. The Mean Squared Error Loss sub-function measures the model's prediction accuracy for continuous numerical values ​​(such as conversion costs), evaluating its regression performance by calculating the squared difference between the predicted and true values. Preset weighting coefficients are parameters used to balance the contribution ratios of the two loss functions to the total loss; adjusting these coefficients allows for prioritizing the optimization of one type of loss, thus achieving a trade-off between different tasks.

[0056] The weighted mixed loss function is a composite loss function that combines multiple loss functions. This loss function design enables the model to optimize classification performance (such as conversion probability prediction) while also taking into account regression performance (such as cost distribution prediction), thereby achieving a unified approach to multi-objective optimization problems.

[0057] The intelligent bidding system constructs a dual-branch machine learning model through the following steps: First, a Long Short-Term Memory (LSTM) temporal branch is established: the LSTM network is configured with two layers, each with 128 neurons. The input dimension is set to temporal features (such as historical click-through rate sequences, conversion rate sequences, and 30-day window data of competitor bidding sequences). The Adam optimizer (learning rate 0.001) is used to train the temporal weight parameters to capture the periodic fluctuations in traffic costs (such as the nighttime cost decline trend), solving the problem that humans cannot respond to dynamic changes in real time. Second, a gradient boosting tree feature branch is established: the maximum depth of the gradient boosting tree is set to 6, and the learning rate is 0.1. The input static features include keyword industry classification (one-hot encoding), audience geographic labels (numerical), age labels (bucket processing), and strategy adjustment parameters (normalized to [-1,1]). A decision tree is constructed using Friedman gradient descent to discover high-conversion combination features such as "first-tier cities + young groups + bidding strategies," overcoming the limitations of human experience.

[0058] Next, a fully connected fusion layer is established: a fully connected layer with 256 neurons is designed, taking into account the 64-dimensional cost trend vector output by LSTM and the 128-dimensional conversion correlation matrix output by GBT. A weighted concatenation method (0.6 weight for the temporal vector, 0.4 weight for the feature matrix) is used for feature fusion, achieving collaborative analysis of spatiotemporal features and business strategies to improve prediction accuracy. Then, a weighted hybrid loss function is constructed: a binary cross-entropy loss (weight 0.6) is used to supervise the conversion probability classification task, and a mean squared error loss (weight 0.4) is used to supervise the cost regression task. The formula is: Total Loss = 0.6 * BCE (actual conversion label, predicted probability) + 0.4 * MSE (actual CPA, predicted cost), where BCE refers to the binary cross-entropy loss and MSE represents the mean squared error loss. Through joint optimization of the two tasks, the accuracy of conversion rate prediction and the stability of cost control are simultaneously improved. Finally, a dual-branch model is integrated: the LSTM temporal branch, the GBT feature branch, the fully connected fusion layer, and the hybrid loss function are encapsulated into a unified model architecture, resulting in a dual-branch machine learning model.

[0059] Step S20: Train the dual-branch machine learning model using historical delivery data to obtain the target prediction model.

[0060] It should be noted that historical campaign data refers to various data records accumulated during past advertising campaigns, including but not limited to keyword clicks, conversion rates, bid amounts, audience characteristics (such as region, age, gender, etc.), ad display time, competitor bidding, and actual conversion costs. Please refer to Table 1 (partial table of historical campaign data fields). The target prediction model refers to a trained and optimized two-branch machine learning model that can predict future advertising performance based on input advertising-related features (such as keywords, audience characteristics, bidding strategies, etc.), such as predicting key indicators like conversion probability and conversion cost.

[0061] Table 1

[0062]

[0063] From this historical delivery data, the model can learn the following association patterns:

[0064] A. Keyword + Time Period: For the keyword "phone installment payment," during 8:00-22:00 (the off-get off work hours in first-tier cities), for every 1 yuan increase in bid, the conversion rate for the keyword "phone installment payment" increases by 8%, but the CPA increases by 12% (the model learns the "conversion potential and cost trade-off during high-traffic periods"). B. Audience + Geographic Pattern: The conversion rate for the keyword "phone installment payment" is 23% higher for the "25-35 years old + first-tier cities" audience than for the "18-24 years old + second-tier cities" audience (the model identifies the conversion difference based on the "age + geographic" combination). C. Competitor Bidding + Own Bidding Pattern: When competitors bid ≥ 10% of your own bid for "fast phone charging," increasing your own bid by 5% can reclaim 15% of traffic, but the CPA will increase by 18% (the model identifies the "bidding game relationship under competitor interference").

[0065] As an example, the historical delivery data includes input data and output labels. The input data includes time-series data and static feature data, and the output labels include conversion rate and cost value. The step of training the dual-branch machine learning model using the historical delivery data to obtain the target prediction model includes: dividing the time-series data into a sliding window according to a preset window length and a preset step size to obtain time-series training samples; performing one-hot encoding and normalization on the static feature data to obtain static feature training samples; inputting the time-series training samples and the static feature training samples into the dual-branch machine learning model to obtain prediction data; calculating a hybrid loss function value based on the prediction data, the conversion rate, and the cost value; and training the dual-branch machine learning model using backpropagation based on the hybrid loss function value to obtain the target prediction model.

[0066] Time-series data refers to serialized data that changes over time. In advertising scenarios, this includes data recorded over time, such as keyword clicks, conversion rates, and competitor bids. Static feature data refers to relatively stable feature data that does not change over time during the advertising process, such as keyword industry classification, audience geographic and age tags, and ad group budgets. Conversion rate refers to the percentage of ad clicks that result in a completed target action (such as purchase or registration), and is an important indicator for measuring advertising effectiveness. Cost per acquisition (CPA) typically refers to the cost per conversion, i.e., the fee paid to achieve a single conversion.

[0067] The preset window length refers to the length of the time period contained in each window when performing a sliding window segmentation on the time series data, for example, 7 days. The preset step size refers to the time step size of each movement of the sliding window, for example, 1 day. These two parameters are used to generate sample sequences for training from the original time series data, ensuring that the model can learn sufficient time-dependent patterns. The time series training samples are sample sequences obtained by performing a sliding window segmentation on the time series data. Each sample contains a time series data segment of fixed length, used for model training. The static feature training samples are static feature data after one-hot encoding and normalization. The predicted data refers to the prediction results generated by the two-branch machine learning model during training based on the input time series training samples and static feature training samples, including estimates of conversion rates and cost values. The mixed loss function value is calculated by comparing the predicted data with the actual conversion rates and cost values, combining the binary cross-entropy loss (used to evaluate the accuracy of conversion rate prediction) and the mean squared error loss (used to evaluate the accuracy of cost value prediction), and weighted summing according to preset weight coefficients.

[0068] First, the intelligent bidding system performs sliding window segmentation on the time-series data according to a preset window length and step size. Specifically, starting from the beginning of the time-series data, the system extracts a segment of data with a length equal to the window length as the first sample, then moves forward according to the step size, extracting the next sample, and repeating this process until all time-series data is covered, thus obtaining a series of time-series training samples. This is done to convert continuous time-series data into fixed-length samples that the model can process, facilitating the model's learning of dynamic patterns in the time series. Second, for categorical features (such as audience geographic tags, keyword industry classifications, etc.), the system uses one-hot encoding to convert them into numerical form, avoiding misjudgments of the numerical magnitude of categorical features by the model; for numerical features (such as ad group budgets, audience age, etc.), the system performs normalization processing, scaling them to the range of 0 to 1 to eliminate dimensional differences between different features, improving the training efficiency and stability of the model.

[0069] Finally, the system inputs both temporal training samples and static feature training samples into a dual-branch machine learning model. The model processes temporal data through the LSTM temporal branch and static feature data through the GBT feature branch, integrating the outputs of both in a fully connected fusion layer to generate predicted data for conversion rates and costs. The system calculates a hybrid loss function based on the difference between the predicted data and the actual conversion rates and costs, and adjusts the model parameters using backpropagation to minimize the loss function value. After multiple iterations of training, a target prediction model capable of accurately predicting advertising performance is obtained. This process ensures that the model can simultaneously learn the dynamic changes in temporal data and the correlation patterns of static features, thus providing accurate predictive support for advertising bidding strategies.

[0070] Step S30: Collect real-time campaign data through Apple Ads marketing interface and generate multiple candidate bidding strategies according to preset business rules.

[0071] It's important to note that real-time delivery data refers to data acquired in real-time during ad delivery via the Apple Ads Marketing API. This includes information such as keyword clicks, conversion rates, impressions, current bids, audience characteristics (e.g., region, age, gender), time-of-day traffic, geographic traffic distribution, and competitor bids. Pre-defined business rules refer to a series of rules and constraints pre-set based on the advertiser's business objectives and strategic needs. These may include target conversion rates, target costs (e.g., CPA), budget constraints, value assessments of different audience groups, and keyword importance and priority.

[0072] Candidate bidding strategies refer to a series of possible bid adjustment schemes generated by the intelligent bidding system based on real-time delivery data and preset business rules. They typically include adjustment step values ​​for keyword bids (such as increasing or decreasing by 10%), audience premium adjustment step values ​​(such as increasing or decreasing the bid for a specific audience group by 20%), time period coefficient adjustment values ​​(such as increasing or decreasing the bid coefficient in a specific time period), and region coefficient adjustment values ​​(such as increasing or decreasing the bid coefficient in a specific region).

[0073] Regarding how to determine high-conversion and low-conversion keywords: 1) Core judgment indicator: The "Conversion Potential Index (TPI)" is used to quantify the value of keywords. TPI is calculated by weighting the following dimensions:

[0074] A. Basic conversion metrics:

[0075] - Historical conversion rate (past 7 days): CTR_7d = Conversions / Clicks;

[0076] -Conversion Cost Stability: CPA Variance (measures cost volatility over the past 30 days);

[0077] B. Traffic Competition Metrics:

[0078] - Competitor bidding intensity: Competitor_Bid = Average competitor bid / Current bid of the competitor;

[0079] - Traffic Popularity Index: Keyword traffic popularity on the Apple Ads platform (0-10 points, obtained from official data interface);

[0080] C. Business objective alignment:

[0081] - Target audience matching: The overlap rate between keywords and the audience tags set by the customer (e.g., the search matching degree between the keyword "mobile phone installment" and the "25-35 years old" audience).

[0082] 2) Calculation formula:

[0083] TPI = (CTR_7d × 0.4) + (1 - CPA variance / mean × 0.2) + (1 - Competitor_Bid × 0.2) + (Traffic_Score × 0.1) + (Audience_Match × 0.1)

[0084] Among them, CTR_7d represents the click-through rate over the past 7 days, which is the ratio of ad clicks to impressions; CPA (Cost Per Action) variance represents the fluctuation range of conversion cost over the past 30 days, while the mean is the average conversion cost over these days; Competitor_Bid represents the bidding intensity of competitors, which is the ratio of the average bid of competitors to the current bid of the advertiser; Traffic_Score is the keyword's traffic popularity index, provided by the Apple Ads platform, which usually ranges from 0 to 10, reflecting the popularity and traffic potential of the keyword; Audience_Match represents the keyword's match with the target audience, which is the overlap rate between the keyword and the audience tags set by the advertiser.

[0085] 3) Threshold determination rules:

[0086] A. Keywords with high conversion potential:

[0087] TPI ≥ industry keyword TPI mean + industry standard deviation × 0.5, and satisfies:

[0088] -Historical conversion rate ≥ industry average × 1.2;

[0089] - Competitor bid strength <1.1 (competitor bids are not significantly higher than our own, there is room for us to gain market share).

[0090] B. Low-conversion keywords:

[0091] TPI ≤ (mean of industry keyword TPI - industry standard deviation × 0.5), or satisfy:

[0092] -Historical conversion rate < industry average × 0.5;

[0093] -CPA variance > industry average × 1.5 (excessive cost fluctuations, high risk);

[0094] - Traffic popularity index <3 (extremely low traffic, poor return on investment).

[0095] 4) Dynamic update mechanism: TPI is recalculated every hour, and the list of high / low conversion keywords is updated daily to adapt to changes in traffic and competitive environment.

[0096] As an example, the step of generating multiple candidate bidding strategies based on preset business rules includes: parsing the preset business rules to obtain a set of step size parameters for keyword adjustment dimension, audience adjustment dimension, time period adjustment dimension, and region adjustment dimension; generating a discrete adjustment value matrix based on the step size parameter set, wherein the keyword adjustment dimension and the audience adjustment dimension generate percentage adjustment values, and the time period adjustment dimension and the region adjustment dimension generate coefficient adjustment values; combining the adjustment values ​​of different dimensions in the discrete adjustment value matrix by Cartesian product to obtain an initial strategy combination set; and filtering the initial strategy combination set according to preset business constraints to obtain multiple candidate bidding strategies.

[0097] Keyword adjustment refers to the scope and direction of bid adjustments made to keyword bids within a bidding strategy. For example, the system might decide to increase bids for certain keywords to increase exposure or decrease bids to save costs, based on their conversion potential and cost-effectiveness. Audience adjustment refers to the scope and direction of bid adjustments made to different audience groups within a bidding strategy. For example, the system might increase bids for high-value audiences (such as young people or users in first-tier cities) to increase the attractiveness of the ads.

[0098] The time-based adjustment dimension refers to the scope and direction of bid adjustments within a bidding strategy for different time periods (such as weekdays, weekends, peak hours, and off-peak hours). For example, increasing bids during periods of lower traffic costs to acquire more traffic, or decreasing bids during periods of higher traffic costs to control costs. The geographic adjustment dimension refers to the scope and direction of bid adjustments within a bidding strategy for different geographic regions (such as first-tier cities, second-tier cities, and lower-tier markets). For example, increasing bids in regions with good conversion rates, or decreasing bids in regions with high traffic costs.

[0099] The step size parameter set refers to the preset adjustment range set for each adjustment dimension. For example, the keyword bid adjustment step size might be ±10% or ±20%, the audience premium adjustment step size might be ±5% or ±10%, and the time period and region adjustment coefficients might be ±0.1 or ±0.2, etc. The discrete adjustment value matrix refers to the specific set of adjustment values ​​generated for each adjustment dimension. It is a multi-dimensional matrix, where each dimension corresponds to an adjustment dimension (such as keyword, audience, time period, region), and the value in each dimension is a discrete adjustment value generated based on the step size parameter set. The percentage adjustment value refers to the bid adjustment value expressed as a percentage in the keyword and audience adjustment dimensions. The coefficient adjustment value refers to the bid adjustment value expressed as a coefficient in the time period and region adjustment dimensions.

[0100] The initial strategy set refers to the set of all possible bidding strategies obtained by combining the Cartesian products of the adjustment values ​​from different dimensions in the discrete adjustment value matrix. Each strategy set includes specific adjustment values ​​for dimensions such as keywords, audience, time period, and region. Pre-set business constraints refer to a series of constraints pre-defined based on the advertiser's business objectives and strategic needs, used to filter out strategies in the initial strategy set that are unreasonable or do not conform to business objectives. For example, pre-set business constraints may include a target CPA (cost per conversion) cap, a target ROI (return on investment) cap, budget limits, and minimum bid requirements for specific audience groups.

[0101] First, the intelligent bidding system parses the preset business rules, extracting step size parameter sets for four adjustment dimensions: keywords, audience, time period, and region. Second, the system generates a discrete adjustment value matrix based on these step size parameter sets. Specifically, for the keyword and audience adjustment dimensions, the system generates a series of percentage adjustment values ​​based on the step size parameters, such as [-20%, -10%, +10%, +20%]; for the time period and region adjustment dimensions, the system generates a series of coefficient adjustment values, such as [-0.2, -0.1, +0.1, +0.2]. These adjustment value matrices provide specific adjustment options for subsequent strategy combinations. Then, the system performs Cartesian product combinations on these adjustment values ​​from different dimensions, that is, it combines each adjustment value in each dimension to generate all possible strategy combinations. For example, if there are 4 options for the keyword adjustment dimension, 3 options for the audience adjustment dimension, and 4 options each for the time period and region adjustment dimensions, then the final generated initial strategy combination set will contain 4 × 3 × 4 × 4 = 192 strategy combinations. Finally, the system filters the initial strategy combination set based on preset business constraints, eliminating those that do not meet business objectives (such as exceeding budget, falling below target conversion rate, click volume > industry 95th percentile, CPA < industry 5th percentile, etc.). It also aggregates the average values ​​of the same keyword + audience combination, thus obtaining multiple candidate bidding strategies that meet the requirements, providing a foundation for subsequent optimization and execution. Industry average calculation:

[0102] Industry average conversion rate = (∑ conversions of advertisers in the same industry) / (∑ clicks of advertisers in the same industry)

[0103] Step S40: Process the real-time delivery data and the candidate bidding strategies using the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies.

[0104] It's important to note that the conversion probability prediction value refers to the likelihood that a click on an ad will convert into actual user behavior (such as purchase, registration, or download) under a specific bidding strategy, as predicted by the target prediction model based on real-time delivery data and candidate bidding strategies. The cost distribution prediction value refers to the cost distribution of ad delivery under a specific bidding strategy, as predicted by the target prediction model based on real-time delivery data and candidate bidding strategies. This includes statistical information such as the mean and variance of costs, representing the possible range and fluctuations of ad delivery costs under that bidding strategy.

[0105] Step S50: Based on the predicted conversion probability and the predicted cost distribution, calculate the strategy score of the candidate bidding strategy using a scoring function, and determine the strategy combination based on the strategy score.

[0106] It should be noted that the scoring function is a mathematical formula or algorithm used to evaluate the overall performance of candidate bidding strategies. It assigns a score to each candidate bidding strategy by comprehensively considering multiple key indicators such as conversion probability prediction and cost distribution prediction. The formula is as follows:

[0107] Score = (P × 0.7) + (1 - C(variance) / 10) × 0.3

[0108] Where P refers to the predicted conversion probability, C refers to the predicted cost distribution, and Score refers to the strategy score.

[0109] A strategy portfolio refers to a set of optimal bidding strategies selected based on strategy score values, such as the top 5 strategies with the highest scores.

[0110] Step S60: Integrate the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters of the strategy combination to generate the optimal bidding strategy package.

[0111] It's important to note that keyword adjustment parameters refer to the specific adjustments made to keyword bids within the final strategy combination. Audience adjustment parameters refer to the bid adjustments made to different audience groups within the strategy combination. Time-of-day adjustment parameters refer to the bid adjustments made to different time periods within the strategy combination; these parameters are expressed as coefficients. Geographic adjustment parameters refer to the bid adjustments made to different regions within the strategy combination; these parameters are also expressed as coefficients.

[0112] An optimal bidding strategy package refers to a complete set of bidding strategies generated by integrating keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters. These strategies have been systematically evaluated and optimized, and are considered to be the bidding solutions most likely to achieve the advertiser's business goals (such as maximizing conversion rates and controlling costs) under current market conditions.

[0113] Bidding constraints: Single adjustment range ≤ 20% (to avoid sudden bid changes); 2 hours before the daily budget is exhausted, the adjustment coefficient is automatically reduced (e.g., from +30% to +10%, extending the campaign period); Real-time verification (if the adjusted bid is greater than the 95th percentile of the industry, a cost warning is triggered, and the adjustment coefficient is automatically rolled back by 50%).

[0114] As an example, the step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package includes: performing dimensional separation processing on the strategy combination to obtain a keyword adjustment set, an audience adjustment set, a time period adjustment set, and a geographic adjustment set; performing weighted aggregation on the keyword adjustment set, calculating a weighted average based on the strategy score value of each strategy in the strategy combination to obtain the target keyword adjustment parameters; selecting the top M audience adjustment parameters with the highest strategy score values ​​from the audience adjustment set; performing trend analysis on the time period adjustment set and the geographic adjustment set, generating a dynamic coefficient curve based on a time series pattern; constructing a structured strategy package based on the target keyword adjustment parameters, the top M audience adjustment parameters, and the dynamic coefficient curve; and encapsulating the structured strategy package using the Apple Ads API protocol to obtain the optimal bidding strategy package.

[0115] The keyword adjustment set refers to the collection of all keyword adjustment parameters separated from the strategy combination, containing the specific adjustment ranges for keyword bids under different strategies. The audience adjustment set refers to the collection of all audience adjustment parameters separated from the strategy combination, containing the bid adjustment ranges for different audience groups under different strategies. The time period adjustment set refers to the collection of all time period adjustment parameters separated from the strategy combination, containing the bid adjustment coefficients for different time periods under different strategies. The geographic adjustment set refers to the collection of all geographic adjustment parameters separated from the strategy combination, containing the bid adjustment coefficients for different geographic regions under different strategies. The target keyword adjustment parameter refers to the final adjustment parameter obtained by weighted aggregation of all adjustment parameters in the keyword adjustment set; it is the specific numerical value of the keyword bid adjustment in the final bidding strategy.

[0116] Basic bidding formula: Target bid = Basic bid × (1 + Strategy adjustment coefficient), where the strategy adjustment coefficient = Keyword adjustment coefficient + Audience adjustment coefficient + Time period adjustment coefficient + Region adjustment coefficient + Competition adjustment coefficient. The calculation logic for each coefficient is as follows:

[0117] A. Keyword Adjustment Coefficient (High / Low Conversion Keywords): High Conversion Keywords: +10% to +30% (The higher the TPI, the larger the coefficient; for example, if the TPI is in the top 10% of the industry, the coefficient is +30%). Low Conversion Keywords: -10% to -50% (The lower the TPI, the smaller the coefficient; for example, if the TPI is in the bottom 5% of the industry, the coefficient is -50%, or even suspension of advertising).

[0118] B. Audience Adjustment Coefficient (High / Low Value Audience): High Value Audience (Historical Conversion Rate > Industry Average × 1.2): +5% to +20% (The higher the audience value, the larger the coefficient; for example, for the "25-35 years old + first-tier city" audience, the coefficient is +20%). Low Value Audience (Historical Conversion Rate < Industry Average × 0.8): -5% to -30% (The lower the audience value, the smaller the coefficient; for example, for the "55 years old + lower-tier market" audience, the coefficient is -30%).

[0119] C. Time Period Adjustment Coefficient (Traffic Cost & Conversion Effect): Low traffic cost + good conversion effect: +5% to +15% (e.g., weekdays 20:00-22:00, coefficient +15%). High traffic cost + poor conversion effect: -5% to -15% (e.g., 2:00-4:00 AM, coefficient -15%).

[0120] D. Regional Adjustment Coefficient (Traffic Cost & Conversion Effect): Low traffic cost + good conversion effect: +3% to +10% (e.g., +10% in first-tier cities on weekdays). High traffic cost + poor conversion effect: -3% to -10% (e.g., -10% in lower-tier markets during the early morning hours).

[0121] E. Competition Adjustment Coefficient (Competitor Bidding Dynamics): Competitor Bid < Your Bid: +0% to +5% (Maintaining competitiveness, coefficient +2%). Competitor Bid > Your Bid × 1.2: +5% to +15% (Seizing traffic, coefficient +10%).

[0122] The dynamic coefficient curve is a time series pattern curve generated by trend analysis of time period adjustment sets and regional adjustment sets. It reflects the changing trend of bid adjustment coefficients over time in different time periods or regions, and can dynamically guide the system to adjust bids according to changes in traffic costs and conversion effects.

[0123] Traffic cost assessment (taking "time-based traffic cost" as an example, the same applies to regions): A. Metric definition: Time-based traffic cost = Total cost of ad placement for that time period / Total traffic (impressions or clicks, configurable); B. Data source: Time-based ad placement data from the Apple Ads platform (hourly / 30-minute granularity, including cost and traffic); historical data + real-time data fusion (historical time-based cost of the past 30 days × 0.7 + real-time time-based cost of the past 7 days × 0.3, smoothing fluctuations). C. Dynamic adjustment: If the traffic cost for a certain time period increases by more than 10% for 3 consecutive days, a traffic cost warning is triggered, and the bid coefficient for that time period is automatically reduced (e.g., from +0.2 to +0.1).

[0124] Conversion Performance Evaluation (taking "Regional Conversion Performance" as an example, the same applies to different time periods): A. Core Metrics: Regional Conversion Rate = Regional Conversion Volume / Regional Click Volume; Regional ROI = (Regional Conversion Revenue - Regional Campaign Cost) / Regional Campaign Cost. B. Data Dimensions: Basic Dimensions: Historical 30-day data (stable period) + Recent 7-day data (recent trend); Enhanced Dimensions: Adjusted based on regional economic level (e.g., per capita GDP data interface) and regional competition intensity (density of competitor campaigns in the same region). C. Dynamic Adjustment: If the conversion rate of a certain region is less than the industry average × 0.8 for two consecutive days, the bid coefficient for that region will be automatically reduced (e.g., from +0.15 to +0.05).

[0125] A structured strategy package is a strategy package that integrates target keyword adjustment parameters, the adjustment parameters of the top M high-scoring audiences, and dynamic coefficient curves. It is a complete structure containing all optimization adjustment parameters, used to guide the advertising system to dynamically adjust bids in a real-time bidding environment.

[0126] First, the intelligent bidding system extracts the keyword adjustment parameters, audience adjustment parameters, time-based adjustment parameters, and geographic adjustment parameters from each strategy, and categorizes them into their corresponding keyword adjustment sets, audience adjustment sets, time-based adjustment sets, and geographic adjustment sets, respectively. Second, the system calculates a weighted average of all keyword adjustment parameters based on the strategy score of each strategy to obtain the target keyword adjustment parameters, ensuring that the final keyword adjustment parameters comprehensively reflect the optimization direction of all strategies. Then, the system selects the top M audience adjustment parameters with the highest strategy scores from the audience adjustment set; this is to identify the audience adjustment parameters that most significantly improve advertising performance. Next, statistical analysis is performed on the time-based and regional adjustment sets. Key indicators such as the average adjustment coefficient, fluctuation range, and trend are calculated for each time period and region. Time series analysis methods (such as moving average, exponential smoothing, or seasonal decomposition) are used to identify periodic patterns and trends in the data. For example, it is found that the adjustment coefficient is generally higher during peak hours on weekdays and lower during off-peak hours on weekends. Based on these analysis results, the system generates a dynamic coefficient curve. This curve, with time on the horizontal axis and the adjustment coefficient on the vertical axis, visually displays the bid adjustment trends for different time periods and regions. Subsequently, the system integrates the target keyword adjustment parameters, the top M audience adjustment parameters, and the dynamic coefficient curve to construct a structured strategy package. This consolidates all optimization adjustment parameters into a complete structure, providing clear parameter settings and adjustment logic for the final bidding strategy. Finally, the structured strategy package is encapsulated using the Apple Ads API protocol, enabling its application in the Apple Ads delivery system to form an optimal bidding strategy package. This allows for dynamic bid optimization in a real-time bidding environment, improving the effectiveness and efficiency of ad delivery.

[0127] As an example, after the step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package, the method further includes: executing the optimal bidding strategy package through the Apple Ads marketing interface and monitoring the actual conversion cost after executing the optimal bidding strategy package; triggering a bidding strategy rollback operation when the actual conversion cost exceeds a preset cost threshold; and sending an abnormal strategy report to a preset ad placement console device when the bidding strategy rollback operation is triggered a preset number of times consecutively.

[0128] Actual conversion cost refers to the average cost paid by the advertiser for each conversion after executing the optimal bidding strategy package. Preset cost threshold refers to the upper limit of cost per conversion pre-set by the advertiser based on business objectives and budget constraints. Preset number of rollbacks refers to the maximum number of rollbacks allowed by the system when bidding strategy rollback operations are triggered consecutively. An abnormal strategy report is a report generated by the system containing detailed information about the abnormal bidding strategy when a pre-set number of bidding strategy rollback operations are triggered consecutively. Preset delivery console device refers to the device used by the advertiser or optimizer to monitor and manage ad delivery, such as a computer or mobile device.

[0129] Please refer to Figure 2 , Figure 2This is a schematic diagram of the strategy rollback process provided in Embodiment 1 of the intelligent bidding method based on the Apple Ads marketing interface of this application. First, the intelligent bidding system synchronizes the keyword, audience, time period, and region adjustment parameters in the optimal bidding strategy package to the advertising delivery system through the Apple Ads marketing interface and starts advertising delivery. The system uses the real-time data feedback interface provided by the Apple Ads platform to continuously monitor the actual conversion cost during the advertising delivery process. Specifically, it obtains the actual conversion cost by collecting the cost data of each conversion and calculating the average value, so as to understand the cost control of advertising delivery in a timely manner. Second, when the monitored actual conversion cost exceeds the preset cost threshold (e.g., when the actual conversion cost (CPA) is > expected cost × 1.15 for three consecutive statistical periods (e.g., every 15 minutes), the system will retrieve the previous verified and stable bidding strategy version from the strategy version library (which must meet the following conditions: CPA fluctuation ≤ 5% in the past 2 hours, conversion volume meets the target), and restore the current keyword bid, audience premium, time period coefficient, and region coefficient to the settings of the stable version. At the same time, the current bidding strategy is paused to prevent further cost overruns. After the rollback is executed, core metrics (CPA, conversion rate, ROI) are immediately verified. If the CPA does not improve within 15 minutes of the rollback, a second rollback is automatically triggered (retrieving an earlier stable version or enabling a "minimum guarantee strategy"). Finally, if the system triggers a preset number of bid strategy rollbacks consecutively, it indicates that there may be a persistent problem with the current bid strategy. The system will automatically generate an abnormal strategy report, detailing the reason for the rollback, the differences between the strategy versions before and after the rollback, and the changes in performance after the rollback. This report will be sent to the preset campaign control console device so that advertisers or optimizers can understand the problem in a timely manner and intervene manually (such as temporarily increasing brand keyword bids, suspending inefficient regional campaigns, and marking the strategy version after intervention as "manually corrected version" and entering an independent monitoring process) to ensure the stability of ad campaigns and cost control.

[0130] In this embodiment, the intelligent bidding system features an API automation execution mechanism. Through the Marketing API interface, it automatically synchronizes generated bidding strategies to the Apple Ads platform, enabling batch updates of bidding parameters (e.g., simultaneously adjusting bids for 1000+ keywords) and supporting custom strategy execution cycles (e.g., automatic optimization every hour). A traffic compensation mechanism automatically releases 5%-10% of the reserve budget for traffic fluctuations caused by rollbacks (e.g., decreased impressions), targeting historically high-conversion periods / regions to quickly restore traffic. Furthermore, the intelligent bidding system is based on a version control mechanism, generating a unique version number for each bidding strategy (e.g., V20250625-1430 represents the strategy at 14:30 on June 25, 2025). When rollback conditions are triggered, the current strategy version is automatically marked as an "abnormal version," and the trigger reason is recorded (e.g., "CPA exceeded expectations by 18.2%").

[0131] As an example, after the step of executing the optimal bidding strategy package through the Apple advertising marketing interface, the method further includes: monitoring the actual conversion probability after executing the optimal bidding strategy package; when the deviation between the actual conversion probability and the predicted conversion probability value is greater than a preset deviation threshold, returning to the step of training the dual-branch machine learning model using historical delivery data to obtain the target prediction model; when the deviation between the actual conversion probability and the predicted conversion probability value is less than or equal to the preset deviation threshold, generating an advertising optimization report based on the optimal bidding strategy package, the actual conversion cost, and the actual conversion probability, including keyword performance comparison dimensions, audience performance comparison dimensions, time period performance comparison dimensions, and regional performance comparison dimensions.

[0132] Actual conversion probability refers to the user conversion rate obtained by monitoring the actual effect of ad placement after implementing the optimal bidding strategy package. Preset deviation threshold refers to the maximum allowable error range between model predictions and actual results. Keyword performance comparison refers to comparing the differences in conversion performance, clicks, and conversion costs for different keywords before and after implementing the optimal bidding strategy package in the ad optimization report. Audience performance comparison refers to comparing the differences in conversion performance, clicks, and conversion costs for different audience groups before and after implementing the optimal bidding strategy package in the ad optimization report. Time period performance comparison refers to comparing the differences in conversion performance, clicks, and conversion costs for different time periods before and after implementing the optimal bidding strategy package in the ad optimization report. Geographic performance comparison refers to comparing the differences in conversion performance, clicks, and conversion costs for different regions before and after implementing the optimal bidding strategy package in the ad optimization report. The ad optimization report is a detailed report generated based on data such as actual conversion costs and actual conversion probabilities after implementing the optimal bidding strategy package.

[0133] First, after executing the optimal bidding strategy package, the system collects ad delivery data in real time through the Apple Ads marketing interface, analyzes conversion behavior after each ad click, and calculates the actual conversion probability, i.e., the ratio of conversions to clicks. Second, the system compares the actual conversion probability with the conversion probability prediction value of the target prediction model, calculating the deviation between the two. If the deviation exceeds a preset deviation threshold, it indicates a significant difference between the model prediction and the actual effect. The system will then re-collect historical delivery data, retrain the dual-branch machine learning model, adjust model parameters to improve prediction accuracy, and ultimately obtain a new target prediction model. Finally, if the deviation is within the preset threshold range, it indicates that the model prediction is relatively accurate. The system will generate an ad optimization report based on the actual performance of the optimal bidding strategy package, combined with the actual conversion cost and actual conversion probability. The report details the changes in performance before and after strategy execution across dimensions such as keywords, audience, time period, and region, providing advertisers with data support to help them evaluate the effectiveness of their bidding strategy and guide subsequent optimization.

[0134] This embodiment provides an intelligent bidding method based on the Apple Ads marketing interface. First, a two-branch machine learning model is constructed, incorporating both a long short-term memory temporal branch and a gradient boosting tree feature branch. This structure fully leverages the advantages of both, improving the model's predictive accuracy and generalization ability. Next, the two-branch machine learning model is trained using historical campaign data. By optimizing the loss function and adjusting the model parameters, a target prediction model capable of accurately predicting conversion probabilities and cost distributions is obtained. Then, real-time campaign data is collected through the Apple Ads marketing interface, and multiple candidate bidding strategies are generated according to preset business rules. Real-time data reflects current market dynamics, providing multiple options for subsequent optimization. Next, the real-time campaign data and candidate bidding strategies are input into the target prediction model. The model outputs the conversion probability prediction and cost distribution prediction values ​​for each candidate bidding strategy, allowing for early evaluation of the effectiveness and cost of different bidding strategies and helping the system select the optimal bidding strategy. Subsequently, based on the conversion probability prediction and cost distribution prediction values, a scoring function is used to calculate the strategy score for each candidate bidding strategy. The strategy combination is determined based on the score value. The scoring function balances conversion probability and cost distribution, ensuring that the selected strategy can both improve conversion rates and control costs. Finally, by integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters in the strategy combination, an optimal bidding strategy package is generated. This strategy package can comprehensively guide the advertising system to dynamically adjust bids in a real-time bidding environment, improve the accuracy and efficiency of advertising, and thus efficiently and automatically optimize Apple Ads advertising bidding strategies to improve campaign efficiency and cost control.

[0135] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the intelligent bidding method based on the Apple Ads marketing interface of this application. The target prediction model includes a long short-term memory temporal branch, a gradient boosting tree feature branch, and a fully connected fusion layer. The real-time delivery data includes temporal features and static features. Step S40 of the intelligent bidding method based on the Apple Ads marketing interface includes steps S41 to S44:

[0136] Step S41: Perform time series encoding on the time series features through the long short-term memory time series branch to obtain the cost trend vector.

[0137] It should be noted that the cost trend vector is a multi-dimensional vector that contains information about the trend of cost changes in time series data, such as the trend of cost increase or decrease, the magnitude of fluctuations, etc.

[0138] Understandably, the system first arranges the temporal features of real-time delivery data (such as keyword clicks, conversion rates, competitor bids, etc.) in chronological order, forming a sequence that is then input into the LSTM temporal branch. Next, the LSTM network processes each data point in the sequence one by one through its internal gating units, dynamically updating the network state using forget gates, input gates, and output gates to capture the temporal dependencies and long-term memory information in the sequence. Finally, after processing by the LSTM network, a fixed-length vector is output. This vector reflects the trend and pattern of cost changes over time, providing a basis for subsequent bidding strategy optimization.

[0139] Step S42: Perform feature combination correlation analysis on the static features and the adjustment parameters in the candidate bidding strategy through the feature branches of the gradient boosting tree to obtain the transformation correlation matrix.

[0140] It should be noted that the conversion correlation matrix is ​​a multi-dimensional matrix that includes the correlation strength between different combinations of features (such as keywords, audience, time period, region, etc.) and conversion probability.

[0141] As an example, the static features include keyword industry classification features, audience geographic tag features, and audience age tag features; the adjustment parameters in the candidate bidding strategy include keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters; the step of performing feature combination association analysis on the static features and the adjustment parameters in the candidate bidding strategy through the gradient boosting tree feature branches to obtain a conversion association matrix includes: performing feature encoding on the keyword industry classification features, the audience geographic tag features, and the audience age tag features to obtain a one-hot encoded vector; and performing feature encoding on the keyword adjustment parameters, the audience adjustment parameters, and the... The time-period adjustment parameters and the region adjustment parameters are normalized to obtain normalized adjustment parameters; the one-hot encoded vector and the normalized adjustment parameters are concatenated to form an enhanced feature vector; the enhanced feature vector is split and decided through a multi-layer decision tree of the gradient boosting tree feature branches to obtain a feature combination rule set; the weight values ​​of the leaf nodes of the multi-layer decision tree are extracted as combination influence factors; a transformation correlation matrix is ​​constructed based on the feature combination rule set and the combination influence factors, where the rows of the transformation correlation matrix represent feature combination categories, the columns of the transformation correlation matrix represent adjustment parameter categories, and the element values ​​of the transformation correlation matrix represent the change in transformation probability.

[0142] Keyword industry classification features refer to the industry category to which the keyword belongs, such as "smartphones," "digital products," and "fashion apparel." Audience geographic tag features refer to the geographical location of the target audience, such as "first-tier cities," "second-tier cities," "lower-tier markets," or specific countries and cities. Audience age tag features refer to the age range of the target audience, such as "18-24 years old," "25-35 years old," and "35 years and older." One-hot encoding vectors are a method of converting categorical features into numerical forms. For each categorical feature, one-hot encoding creates a binary vector with only one element set to 1 and the rest to 0, representing the category of that feature.

[0143] Normalization adjustment parameters refer to the process of scaling adjustment parameters to a uniform range (usually 0 to 1). Enhanced feature vectors are feature vectors formed by concatenating one-hot encoded vectors and normalized adjustment parameters, combining information from categorical and numerical features. Feature combination rule sets are a set of rules obtained by splitting the enhanced feature vectors through a multi-layer decision tree with gradient boosting tree feature branches. These rules describe the relationships between different feature combinations, such as "if the keyword belongs to the 'smartphone' industry and the audience age is '25-35 years old,' then the conversion probability increases by 10%." Combination influence factors are weight values ​​extracted from the leaf nodes of a multi-layer decision tree, used to represent the degree of influence of feature combinations on conversion probability.

[0144] First, create a one-hot encoding vector for each categorical feature. For example, if there are three keyword industry categories (A, B, C), encode them as [1,0,0], [0,1,0], and [0,0,1] respectively. This converts the categorical features into numerical form, making them easier for the model to process. Second, scale the values ​​of the keyword adjustment parameter, audience adjustment parameter, time period adjustment parameter, and region adjustment parameter to the range of 0 to 1. This eliminates the dimensional differences between different parameters, improving the efficiency and stability of model training. Then, concatenate the one-hot encoding vector with the normalized adjustment parameter to form an enhanced feature vector. For example, if a keyword belongs to industry A, the audience region is first-tier cities, and the keyword adjustment parameter is +10%, the enhanced feature vector might be [1,0,0,1,0,0,0.5] (assuming the adjustment parameter is 0.5). This integrates the information from categorical and numerical features into a single vector.

[0145] Next, the enhanced feature vectors are input into the top-level decision tree node of the gradient boosting tree's feature branch. The decision tree divides the data according to preset splitting rules (usually based on threshold judgments of feature values). For example, if the value of a feature is greater than a set threshold, the data enters the left subtree; otherwise, it enters the right subtree. This process is recursively performed layer by layer. Each layer of the decision tree node splits according to different features and thresholds until a preset tree depth or node purity requirement is met, ultimately forming a multi-layered decision tree structure. In this way, the model can learn the complex relationships between different feature combinations. Subsequently, the weight values ​​of the leaf nodes of the multi-layered decision trees are extracted as combination influence factors. If the weight value of a leaf node is high, it indicates that the feature combination has a greater impact on the conversion probability. Finally, a conversion correlation matrix is ​​constructed based on the feature combination rule set and the combination influence factors. The rows of the matrix represent the feature combination category, the columns represent the adjustment parameter category, and the element values ​​of the matrix represent the change in conversion probability under a specific feature combination and adjustment parameter. For example, a value of 0.1 in the matrix indicates that, under a specific combination of keyword industry category and audience geographic tag, when the keyword adjustment parameter is +10%, the conversion probability increases by 10%. This matrix can intuitively show the impact of different feature combinations on conversion probability, providing an important reference for optimizing bidding strategies.

[0146] Step S43: The cost trend vector and the conversion correlation matrix are fused through the fully connected fusion layer to obtain joint prediction features.

[0147] Understandably, in the fully connected fusion layer, the cost trend vector and conversion correlation matrix are integrated through a series of linear transformations and non-linear activation functions (such as ReLU). Specifically, the cost trend vector and conversion correlation matrix are flattened into one-dimensional vectors and then concatenated to form a longer feature vector. This concatenated feature vector is then processed through one or more fully connected neural network layers. Each layer performs weighted summation and non-linear transformations on the input features, progressively extracting higher-level feature representations. Finally, after processing by the fully connected fusion layer, a joint predictive feature vector that integrates cost trend and conversion correlation information is output. This joint predictive feature vector can simultaneously reflect the cost dynamics and conversion potential of advertising, providing more comprehensive and accurate information support for subsequent bidding strategy predictions.

[0148] Step S44: Perform probability distribution transformation and cost quantile mapping on the joint prediction features to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy.

[0149] Understandably, the first step is to map the joint prediction features to the [0,1] interval using a probability transformation function (such as the sigmoid function) to obtain the conversion probability prediction value corresponding to each candidate bidding strategy. This converts the raw feature values ​​output by the model into probabilistic values, facilitating subsequent decision-making and evaluation. Then, based on the quantiles of the cost distribution in historical data, the cost-related components of the joint prediction features are mapped to the actual cost interval. For example, if the 25th quantile of historical cost data is 8 yuan and the 75th quantile is 12 yuan, the cost value predicted by the model will be mapped to this interval, obtaining the cost distribution prediction value for each candidate bidding strategy. In this way, the system can align the model's output with the actual cost and conversion probability distribution, thus providing a more accurate predictive basis for bidding strategy selection.

[0150] The following is a table of module functions for the intelligent bidding system:

[0151] Table 2

[0152]

[0153]

[0154] Please refer to Figure 4 , Figure 4This diagram illustrates the core module interactions of the intelligent bidding method based on the Apple Ads marketing interface, as described in Embodiment 2 of this application. The system begins with an Apple Ads account or campaign requirements, acquiring real-time data through a data acquisition module. The model training module then uses this data to generate a conversion prediction model. Next, the strategy generation module calculates multi-dimensional bidding strategies based on the prediction model. A risk control module validates these strategies, assessing bid fluctuations and costs to ensure feasibility and security. If a strategy passes validation, a version control / compensation validation module further confirms it to ensure stability. If validation fails, an anomaly report / intervention interface is triggered for manual intervention. Once a strategy is confirmed, an automated execution module synchronizes it to the Apple Ads platform. An effectiveness monitoring module tracks campaign data and evaluates strategy performance. If performance targets are met, the system generates a campaign report and updates the platform; if not, an iterative optimization module adjusts the model or strategy to continuously improve the bidding strategy. The entire process aims to achieve automated, data-driven optimization of advertising bidding strategies to improve advertising efficiency and ROI.

[0155] This embodiment first uses an LSTM temporal branch to encode the temporal features, capturing dynamic changes and long-term dependencies in the time-series data to obtain a cost trend vector, providing a time-dimensional reference for subsequent bidding strategy optimization. Next, a GBT feature branch is used to perform feature combination correlation analysis on static features and adjustment parameters in candidate bidding strategies, obtaining a conversion correlation matrix. This step visually demonstrates the contribution of different feature combinations to conversion probability, providing data support for bidding strategy adjustments. Then, a fully connected fusion layer fuses the cost trend vector and conversion correlation matrix to generate joint prediction features, comprehensively considering information from both time and feature combination dimensions to generate a more comprehensive feature representation. Finally, probability distribution transformation and cost quantile mapping are performed on the joint prediction features to obtain the predicted conversion probability and cost distribution values ​​of the candidate bidding strategies. This step aligns the model's output with actual business needs, ensuring the prediction results have practical significance. The entire process, by comprehensively considering multi-dimensional information, improves the scientific rigor and accuracy of bidding strategies, achieving automated optimization of ad placement.

[0156] For example, to help understand the implementation process of the smart bidding method based on the Apple Ads marketing interface obtained by combining this embodiment with the above embodiment one, please refer to... Figure 5 , Figure 5 A simplified flowchart illustrating a smart bidding method based on Apple's advertising marketing interface is provided, specifically:

[0157] The process begins with initializing the system and deployment parameters, using a data acquisition module to pull Apple Ads data in real time. Next, the model training module uses this data to generate a conversion prediction model, while the strategy generation module calculates multi-dimensional bidding strategies based on this model. The risk control module is responsible for verifying bid fluctuations and costs to ensure the stability and cost-effectiveness of the strategy. If the verification fails, the system triggers a strategy rollback or manual review; if the verification passes, version control is implemented, reverting to a stable strategy. Then, the automated execution module synchronizes the strategy to the platform, and the performance monitoring module tracks deployment data to evaluate whether the performance targets are met. If the performance targets are met, the system outputs an optimization report; if not, it enters the iterative optimization module to adjust the model or strategy, then loops back to the data acquisition stage to continuously optimize the bidding strategy. The entire process aims to efficiently and automatically optimize Apple Ads advertising bidding strategies through automation and data-driven methods, thereby improving deployment efficiency and cost control.

[0158] This application also provides a smart bidding device based on Apple's advertising marketing interface, please refer to... Figure 6 The smart bidding device based on the Apple Ads marketing interface includes:

[0159] Model building module 10 is used to build a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch;

[0160] Model training module 20 is used to train the dual-branch machine learning model using historical deployment data to obtain a target prediction model;

[0161] The data acquisition module 30 is used to collect real-time delivery data through the Apple Ads marketing interface and generate multiple candidate bidding strategies according to preset business rules.

[0162] Prediction module 40 is used to process the real-time delivery data and the candidate bidding strategy through the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy;

[0163] The strategy scoring module 50 is used to calculate the strategy score of the candidate bidding strategy based on the conversion probability prediction value and the cost distribution prediction value through a scoring function, and to determine the strategy combination based on the strategy score value;

[0164] The strategy determination module 60 is used to integrate the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters of the strategy combination to generate the optimal bidding strategy package.

[0165] This application provides a smart bidding device based on the Apple Ads Marketing Interface. The smart bidding device based on the Apple Ads Marketing Interface includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the smart bidding method based on the Apple Ads Marketing Interface in the above embodiment 1.

[0166] This application provides a computer-readable storage medium having computer-readable program instructions stored thereon, the computer-readable program instructions being used to execute the smart bidding method based on the Apple Ads Marketing Interface in the above embodiments.

[0167] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the smart bidding method based on the Apple Ads Marketing Interface described above.

[0168] The intelligent bidding device, equipment, storage medium, and computer program product based on the Apple Ads marketing interface provided in this application, employing the intelligent bidding method based on the Apple Ads marketing interface in the above embodiments, can solve the technical problem of how to efficiently and automatically optimize Apple Ads advertising bidding strategies to improve campaign efficiency and cost control. Compared with the prior art, the beneficial effects of the intelligent bidding device, equipment, storage medium, and computer program product based on the Apple Ads marketing interface provided in this application are the same as the beneficial effects of the intelligent bidding method based on the Apple Ads marketing interface provided in the above embodiments, and other technical features are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0169] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A smart bidding method based on Apple's advertising marketing interface, characterized in that, The method includes: Construct a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch; The dual-branch machine learning model is trained using historical delivery data to obtain a target prediction model; Real-time campaign data is collected through Apple Ads marketing API, and multiple candidate bidding strategies are generated based on preset business rules; The real-time delivery data and the candidate bidding strategies are processed by the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies. Based on the predicted conversion probability and the predicted cost distribution, the strategy score of the candidate bidding strategy is calculated using a scoring function, and the strategy combination is determined based on the strategy score. By integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters of the aforementioned strategy combination, an optimal bidding strategy package is generated. The target prediction model includes a long short-term memory temporal branch, a gradient boosting tree feature branch, and a fully connected fusion layer; the real-time delivery data includes temporal features and static features. The step of processing the real-time delivery data and the candidate bidding strategies using the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies includes: The time-series features are encoded using the long short-term memory time-series branch to obtain a cost trend vector. By performing feature combination correlation analysis on the static features and the adjustment parameters in the candidate bidding strategy through the feature branches of the gradient boosting tree, a conversion correlation matrix is ​​obtained; The cost trend vector and the conversion correlation matrix are fused using the fully connected fusion layer to obtain joint prediction features. The joint prediction features are subjected to probability distribution transformation and cost quantile mapping to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy.

2. The method as described in claim 1, characterized in that, The steps for constructing a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch include: A long short-term memory time-series branch is established, which is used to process time-series data containing keyword click-through rate sequences, conversion rate sequences, and competitor bidding sequences. The long short-term memory time-series branch uses an adaptive moment estimation optimizer to update parameters. A gradient boosting tree feature branch is established, which is used to process static feature data containing keyword industry classification features, audience geographic tag features, audience age tag features, and ad group budget features. The gradient boosting tree feature branch is constructed using the Friedman gradient descent optimizer. A fully connected fusion layer is established to connect the output nodes of the long short-term memory temporal branch and the output nodes of the gradient boosting tree feature branch; A weighted mixed loss function is constructed based on the binary cross-entropy loss function, the mean squared error loss function, and the preset weight coefficients. By integrating the long short-term memory temporal branch, the gradient boosting tree feature branch, the fully connected fusion layer, and the weighted hybrid loss function, a two-branch machine learning model is obtained.

3. The method as described in claim 1, characterized in that, The static features include keyword industry classification features, audience geographic tag features, and audience age tag features. The adjustment parameters in the candidate bidding strategy include keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters. The step of performing feature combination correlation analysis on the static features and the adjustment parameters in the candidate bidding strategy through the feature branches of the gradient boosting tree to obtain the transformation correlation matrix includes: The keyword industry classification features, the audience geographic tag features, and the audience age tag features are feature-encoded to obtain a one-hot encoded vector; The keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters are normalized to obtain normalized adjustment parameters. The one-hot encoded vector is concatenated with the normalization adjustment parameter to form an enhanced feature vector; The enhanced feature vector is split and decision-making process is performed on the multi-layer decision tree of the gradient boosting tree feature branches to obtain a set of feature combination rules. The weight values ​​of the leaf nodes of the multi-layer decision tree are extracted as combined influence factors; A transformation correlation matrix is ​​constructed based on the set of feature combination rules and the combination influence factors. The rows of the transformation correlation matrix represent the feature combination categories, the columns of the transformation correlation matrix represent the adjustment parameter categories, and the element values ​​of the transformation correlation matrix represent the change in transformation probability.

4. The method as described in claim 1, characterized in that, The step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package includes: The strategy combination is subjected to dimensional separation processing to obtain keyword adjustment set, audience adjustment set, time period adjustment set and region adjustment set; The keyword adjustment set is weighted and aggregated, and the weighted average value is calculated based on the strategy score of each strategy in the strategy combination to obtain the target keyword adjustment parameters; Select the top M audience adjustment parameters with the highest strategy score from the audience adjustment set; Trend analysis is performed on the time-period adjustment set and the regional adjustment set, and a dynamic coefficient curve is generated based on the time series pattern; A structured strategy package is constructed based on the target keyword adjustment parameters, the top M audience adjustment parameters, and the dynamic coefficient curve; The structured strategy package is encapsulated using the Apple Ads API protocol to obtain the optimal bidding strategy package.

5. The method as described in claim 1, characterized in that, The step of generating multiple candidate bidding strategies according to preset business rules includes: The preset business rules are parsed to obtain a set of step size parameters for keyword adjustment dimension, audience adjustment dimension, time period adjustment dimension, and region adjustment dimension; A discrete adjustment value matrix is ​​generated based on the set of step size parameters, wherein the keyword adjustment dimension and the audience adjustment dimension generate percentage adjustment values, and the time period adjustment dimension and the region adjustment dimension generate coefficient adjustment values; The Cartesian product of the adjustment values ​​in different dimensions of the discrete adjustment value matrix is ​​combined to obtain the initial strategy combination set; The initial strategy combination set is filtered according to preset business constraints to obtain multiple candidate bidding strategies.

6. The method as described in claim 1, characterized in that, The historical delivery data includes input data and output tags. The input data includes time-series data and static feature data. The output tags include conversion rate and cost value. The step of training the dual-branch machine learning model using historical delivery data to obtain the target prediction model includes: The time series data is divided into time series training samples by sliding window segmentation according to preset window length and preset step size; The static feature data is one-hot encoded and normalized to obtain static feature training samples; The time-series training samples and the static feature training samples are input into the dual-branch machine learning model to obtain prediction data; Calculate the hybrid loss function value based on the predicted data, the conversion rate, and the cost value; Based on the hybrid loss function value, the dual-branch machine learning model is trained using backpropagation to obtain the target prediction model.

7. The method as described in claim 1, characterized in that, After the step of integrating the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and geographic adjustment parameters of the strategy combination to generate the optimal bidding strategy package, the method further includes: The optimal bidding strategy package is executed through the Apple Ads marketing interface, and the actual conversion cost after executing the optimal bidding strategy package is monitored. If the actual conversion cost exceeds a preset cost threshold, a bid strategy rollback operation is triggered. If the bidding strategy rollback operation is triggered a preset number of times consecutively, an abnormal strategy report is sent to the preset delivery console device.

8. The method as described in claim 7, characterized in that, After the step of executing the optimal bidding strategy package through the Apple Ads marketing interface, the method further includes: Monitor the actual conversion probability after executing the optimal bidding strategy package; If the deviation between the actual conversion probability and the predicted conversion probability is greater than a preset deviation threshold, return to the step of training the dual-branch machine learning model using historical delivery data to obtain the target prediction model. When the deviation between the actual conversion probability and the predicted conversion probability is less than or equal to the preset deviation threshold, an advertising optimization report is generated based on the optimal bidding strategy package, the actual conversion cost, and the actual conversion probability, including keyword performance comparison, audience performance comparison, time period performance comparison, and regional performance comparison.

9. A smart bidding device based on Apple's advertising marketing interface, characterized in that, The device includes: The model building module is used to build a two-branch machine learning model that includes a long short-term memory temporal branch and a gradient boosting tree feature branch. The model training module is used to train the dual-branch machine learning model using historical deployment data to obtain the target prediction model; The data acquisition module is used to collect real-time campaign data through the Apple Ads marketing interface and generate multiple candidate bidding strategies based on preset business rules. The prediction module is used to process the real-time delivery data and the candidate bidding strategies through the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies. The target prediction model includes a long short-term memory time-series branch, a gradient boosting tree feature branch, and a fully connected fusion layer. The real-time delivery data includes time-series features and static features. The step of processing the real-time delivery data and the candidate bidding strategies through the target prediction model to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategies includes: performing time-series encoding processing on the time-series features through the long short-term memory time-series branch to obtain a cost trend vector; performing feature combination correlation analysis on the static features and the adjustment parameters in the candidate bidding strategies through the gradient boosting tree feature branch to obtain a conversion correlation matrix; performing feature fusion on the cost trend vector and the conversion correlation matrix through the fully connected fusion layer to obtain a joint prediction feature; and performing probability distribution transformation and cost quantile mapping on the joint prediction feature to obtain the conversion probability prediction value and cost distribution prediction value of the candidate bidding strategy. The strategy scoring module is used to calculate the strategy score of the candidate bidding strategy based on the conversion probability prediction value and the cost distribution prediction value through a scoring function, and to determine the strategy combination based on the strategy score value; The strategy determination module is used to integrate the keyword adjustment parameters, audience adjustment parameters, time period adjustment parameters, and region adjustment parameters of the strategy combination to generate the optimal bidding strategy package.

Images