Marketing data synthesis and augmentation method and system based on generative adversarial network

Abstract

The application relates to the technical field of digital marketing data processing, and in particular to a marketing data synthesis and enhancement method and system based on a generative adversarial network. The method comprises the following steps: acquiring original marketing data, extracting multi-modal features, and inputting the multi-modal features into a generative adversarial network comprising a generator, a reality discriminator and an effect discriminator; using the effect discriminator to evaluate the probability of the synthesized data producing a positive marketing effect and generating a signal; calculating a joint loss function of the generator and dynamically adjusting a weight coefficient, and updating the generator parameters according to the joint loss function and the weight coefficient to optimize the quality of the synthesized data; and finally mixing the optimized synthesized data and the original data to form an enhanced data set. The application realizes the automation and intelligentization of marketing data synthesis, significantly improves the marketing effect while guaranteeing the data authenticity, and effectively protects the data privacy and security.

Description

Technical Field

[0001] This application relates to the technical field of digital marketing data processing, and in particular to a method and system for marketing data synthesis and enhancement based on generative adversarial networks. Background Technology

[0002] With the rapid development of the digital marketing industry and the continuous growth in demand for personalized marketing, enterprises need to process massive amounts of marketing data on a daily basis to optimize ad placement, build user profiles, and formulate marketing strategies. This data covers multi-dimensional information such as user behavior patterns, consumption preferences, and multi-channel interaction records, and is characterized by its large volume, rapid updates, and complex structure.

[0003] Currently, traditional marketing data augmentation methods mainly rely on simple oversampling or interpolation techniques, such as SMOTE (Synthetic Minority Oversampling), which generates new data by linearly combining existing samples. This approach is effective when dealing with low-dimensional and uniformly distributed marketing data. However, when faced with high-dimensional, sparse, and dynamically changing real-world marketing scenarios (such as user clickstream data or cross-channel conversion paths), the synthetic data generated by traditional methods often lacks authenticity and diversity, making it difficult to simulate complex user behavior patterns. This leads to overfitting or insufficient generalization ability in downstream machine learning models (such as click-through rate prediction models), directly affecting the accuracy of marketing decisions.

[0004] While existing data synthesis techniques partially employ Generative Adversarial Networks (GAN) frameworks, they largely focus on generating single-modal data such as images and text, lacking targeted optimization for multimodal data specific to marketing scenarios (such as user profiles, temporal behavior, and contextual features). Furthermore, conventional GANs prioritize the realism of generated data, with their discriminators only evaluating the distributional differences between generated and real samples, failing to incorporate marketing effectiveness-oriented evaluation mechanisms. This makes it impossible to ensure the effectiveness and operability of synthesized data for specific marketing goals (such as improving conversion rates and optimizing user retention). For example, in ad placement optimization, while synthesized data may appear realistic, it may fail to drive actual user conversion behavior, resulting in a waste of marketing resources.

[0005] Furthermore, the current marketing data synthesis process suffers from weak privacy protection mechanisms. During training, the generative model may memorize sensitive personal information from the original data (such as user identification and purchase records). If the synthesized data is directly output for third-party analysis or sharing, there is a risk of privacy breaches.

[0006] Therefore, existing technologies lack an intelligent data synthesis and enhancement solution that can simultaneously meet the requirements of multimodal data fusion, marketing effectiveness orientation, and privacy protection. There is an urgent need to improve the automation, accuracy, and security of marketing data processing through innovative methods. Summary of the Invention

[0007] To automate and intelligently synthesize marketing data, significantly improve marketing effectiveness while ensuring data authenticity, and effectively protect data privacy and security, this application provides a marketing data synthesis and enhancement method, system, device, and medium based on generative adversarial networks.

[0008] The above-mentioned objective of this application is achieved through the following technical solution:

[0009] A marketing data synthesis and enhancement method based on generative adversarial networks includes the following steps:

[0010] Obtain the original marketing dataset and extract multimodal marketing features based on the original marketing dataset;

[0011] The multimodal marketing features are input into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, an authenticity discriminator, and an effect discriminator;

[0012] The effect discriminator is used to evaluate the probability that the synthetic marketing data output by the generator will produce a positive marketing effect in a predetermined marketing scenario, and an effect evaluation signal is generated.

[0013] Based on the output of the authenticity discriminator and the effect evaluation signal, the joint loss function of the generator is calculated, and the weight coefficients of the authenticity loss term and the effect loss term in the joint loss function are dynamically adjusted.

[0014] The generator parameters are updated according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effectiveness.

[0015] The optimized synthetic marketing data is then mixed with the original marketing dataset to form an enhanced marketing dataset.

[0016] By adopting the above technical solution, and by acquiring the original marketing dataset and extracting multimodal marketing features, this method automatically integrates multi-source data such as user profiles and behavioral sequences, solving the sparsity and heterogeneity problems of traditional marketing data, improving the comprehensiveness and accuracy of data representation, and providing high-quality input for subsequent generative adversarial networks, thereby reducing manual processing costs and strengthening the data foundation. The multimodal features are input into an effect-oriented generative adversarial network. Through a three-element structure of generator, realism discriminator, and effect discriminator, marketing objectives are directly integrated into the training process, ensuring that the synthesized data is not only realistic but also has business value, significantly improving the targeting and efficiency of marketing activities. The effect discriminator is used to evaluate the probability of the synthesized data generating positive marketing effects, generating effect evaluation signals, achieving real-time feedback and quantitative evaluation, enabling the generator to dynamically adjust its output, reducing the uncertainty of marketing activities and improving the return on investment. Based on the output of the realism discriminator... By calculating a joint loss function with the effect evaluation signal and dynamically adjusting the weight coefficients, the model performance is adaptively optimized by balancing the loss of realism and the loss of effect, avoiding overfitting or underfitting, and enhancing the robustness and scenario adaptability of the synthesized data. The generator parameters are updated according to the joint loss function to optimize the quality of the synthesized data, realizing end-to-end automated training, ensuring the consistency between data realism and effect orientation, and improving generation efficiency and reliability. The optimized synthesized data is mixed with the original dataset to form an augmented dataset. The sample size is expanded through a simple and efficient mixing strategy, introducing diversity and solving the problem of poor model generalization ability caused by insufficient data. This supports the accuracy and robustness of downstream marketing models, reduces operating costs, and improves the feasibility of large-scale personalized marketing. It not only improves the intelligence and accuracy of marketing data synthesis, but also ensures long-term practicality through a closed-loop optimization mechanism, providing an efficient and reliable solution for the digital marketing field.

[0017] In a preferred example, this application can be further configured as follows: obtaining the original marketing dataset and extracting multimodal marketing features based on the original marketing dataset specifically includes:

[0018] Obtain a raw marketing dataset containing at least two of the following: user profile data, user behavior sequence data, product information data, and marketing context data;

[0019] The original marketing dataset is cleaned, missing values ​​are removed, and outliers are detected to obtain a standardized marketing data sequence.

[0020] Based on the processed marketing data sequence, multimodal marketing data features are extracted. An attention mechanism is used to perform cross-modal alignment and fusion of the extracted multimodal marketing data features to generate a unified multimodal marketing feature vector. Each feature vector sample is labeled with its corresponding marketing funnel stage label.

[0021] By adopting the above technical solutions, data preprocessing and feature fusion eliminate noise and inconsistencies in the original data through cleaning, alignment, and attention mechanisms, generating unified multimodal feature vectors, improving the accuracy and efficiency of feature extraction, enhancing the semantic integrity of the data through cross-modal alignment, providing structured high-quality input for generative adversarial networks, reducing error accumulation in subsequent steps, and improving the overall system stability.

[0022] In a preferred embodiment, this application can be further configured such that the effect-oriented generative adversarial network is constructed and its input specifically includes:

[0023] The generator is constructed with inputs including a random noise vector and a marketing target vector, wherein the marketing target vector is used to represent the desired marketing objective, including the target user group identifier, the type of marketing metric to be improved, and the target improvement amount;

[0024] The authenticity discriminator is constructed to receive real marketing data samples or synthetic marketing data samples generated by the generator, and output the probability that the sample is a real sample.

[0025] The effect discriminator is constructed based on historical marketing campaign data, which includes positive and negative samples. The effect discriminator is used to evaluate the predictive probability that the input data sample can lead to the target user behavior in a given marketing scenario.

[0026] The multimodal marketing feature vector is combined with the corresponding marketing target vector and then input into the generative adversarial network for training.

[0027] By adopting the above technical solutions, the construction of the generator, authenticity discriminator, and effect discriminator is modularly designed, achieving seamless integration of marketing target vectors. This makes the generation process controllable and goal-driven. Training the effect discriminator with historical data ensures the objectivity of the evaluation. At the same time, the input combination optimizes the network training efficiency, improves the alignment between synthetic data and business needs, and enhances the practicality and scalability of the method.

[0028] In a preferred embodiment, this application can be further configured to utilize the effect discriminator to evaluate the probability that the synthetic marketing data output by the generator will produce a positive marketing effect in a predetermined marketing scenario, and to generate an effect evaluation signal, specifically including:

[0029] The synthetic marketing data sample generated by the generator based on the random noise vector and the marketing target vector is input into the effect discriminator;

[0030] The effect discriminator calculates the conditional probability of the synthetic marketing data sample corresponding to the target user behavior specified in the marketing objective vector based on its internal model parameters, and uses this probability as the probability of the positive marketing effect;

[0031] The calculated conditional probability is compared with a preset effect threshold to generate a binarized preliminary effect evaluation signal.

[0032] Based on the continuous value of the conditional probability or the binarized preliminary effect evaluation signal, combined with the target improvement magnitude in the marketing target vector, the final effect evaluation signal is generated through a predefined mapping function.

[0033] By adopting the above technical solution, the effect evaluation signal is generated through probability calculation, threshold comparison and mapping function, realizing the transformation from quantitative evaluation to decision signal, providing a clear feedback mechanism, enabling the generator to respond quickly to changes in marketing scenarios, reducing the risk of false alarms, and supporting fine-grained optimization through continuous value processing, thereby improving the adaptability and effect orientation of the synthesized data.

[0034] In a preferred embodiment, this application can be further configured as follows: Based on the output of the realism discriminator and the effect evaluation signal, the joint loss function of the generator is calculated, and the weight coefficients of the realism loss term and the effect loss term in the joint loss function are dynamically adjusted, specifically including:

[0035] Calculate the authenticity loss term of the generator relative to the authenticity discriminator, the authenticity loss term being calculated based on the discrimination result of the authenticity discriminator on the synthetic marketing data;

[0036] Calculate the effect loss term of the generator relative to the effect discriminator, the effect loss term being calculated based on the effect evaluation signal;

[0037] Based on the current training stage, the sparsity of the original marketing dataset, or the target improvement specified in the marketing target vector, the weight coefficients α and β of the authenticity loss term and the effect loss term in the joint loss function are dynamically determined.

[0038] The overall joint loss function of the generator is calculated using dynamically adjusted weighting coefficients α and β.

[0039] By adopting the above technical solutions, dynamic weight adjustment is based on training context and data features. Loss balance is optimized in real time through heuristic rules or optimization algorithms, ensuring the best performance of the model in different scenarios and avoiding bias caused by a single objective. This improves the overall quality of synthetic data and supports the flexible implementation of complex marketing strategies.

[0040] In a preferred embodiment, this application may be further configured to: after mixing the optimized synthetic marketing data with the original marketing dataset to form an enhanced marketing dataset, further comprising:

[0041] Based on the enhanced marketing dataset, obtain marketing performance feedback data;

[0042] The marketing performance feedback data is fed back to the performance discriminator to update its parameters, forming a closed-loop optimization system.

[0043] By adopting the above technical solutions, the feedback of marketing performance data is continuously updated and the model is iterated through a closed-loop optimization mechanism. Through real-time data injection, the system can adapt to dynamic market changes, improve long-term performance and maintainability, reduce the need for manual parameter tuning, and enhance the self-learning ability and sustainability of the method.

[0044] In a preferred embodiment, this application can be further configured such that the marketing data synthesis and enhancement method based on generative adversarial networks also includes:

[0045] Before inputting raw marketing data into the feature extraction process, direct identifiers are deleted or desensitized.

[0046] During the training process of the generator, noise that meets the differential privacy requirements is added to the random noise vector input to it, and the scale of the noise is adaptively adjusted according to a preset privacy budget.

[0047] The synthesized data output by the generator is checked, and the synthesized marketing data is only used for subsequent enhancement and application steps after it meets privacy and security standards.

[0048] By adopting the above technical solutions, privacy protection, through desensitization, differential noise reduction, and output inspection, constructs an end-to-end protection system, ensuring that the data synthesis process complies with regulatory requirements. Through quantifiable privacy budgets and adaptive adjustments, it balances data utility and security, reduces the risk of information leakage, enhances the system's credibility and compliance, and provides a secure foundation for marketing data applications.

[0049] The second objective of this invention is achieved through the following technical solution:

[0050] A marketing data synthesis and enhancement system based on generative adversarial networks includes:

[0051] Marketing data processing module: used to acquire the original marketing dataset and extract multimodal marketing features based on the original marketing dataset;

[0052] Generative Adversarial Network Training Module: Used to input the multimodal marketing features into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, a realism discriminator, and an effect discriminator;

[0053] Marketing effectiveness evaluation module: used to evaluate the probability that the synthetic marketing data output by the generator will produce positive marketing effects in a predetermined marketing scenario using the effectiveness discriminator, and generate an effectiveness evaluation signal;

[0054] Dynamic adjustment module: used to calculate the joint loss function of the generator based on the output of the realism discriminator and the effect evaluation signal, and dynamically adjust the weight coefficients of the realism loss term and the effect loss term in the joint loss function;

[0055] The data optimization module is used to update the parameters of the generator according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effect orientation.

[0056] The marketing data enhancement module is used to mix the optimized synthetic marketing data with the original marketing dataset to form an enhanced marketing dataset.

[0057] By adopting the above technical solution, and by acquiring the original marketing dataset and extracting multimodal marketing features, this method automatically integrates multi-source data such as user profiles and behavioral sequences, solving the sparsity and heterogeneity problems of traditional marketing data, improving the comprehensiveness and accuracy of data representation, and providing high-quality input for subsequent generative adversarial networks, thereby reducing manual processing costs and strengthening the data foundation. The multimodal features are input into an effect-oriented generative adversarial network. Through a three-element structure of generator, realism discriminator, and effect discriminator, marketing objectives are directly integrated into the training process, ensuring that the synthesized data is not only realistic but also has business value, significantly improving the targeting and efficiency of marketing activities. The effect discriminator is used to evaluate the probability of the synthesized data generating positive marketing effects, generating effect evaluation signals, achieving real-time feedback and quantitative evaluation, enabling the generator to dynamically adjust its output, reducing the uncertainty of marketing activities and improving the return on investment. Based on the output of the realism discriminator... By calculating a joint loss function with the effect evaluation signal and dynamically adjusting the weight coefficients, the model performance is adaptively optimized by balancing the loss of realism and the loss of effect, avoiding overfitting or underfitting, and enhancing the robustness and scenario adaptability of the synthesized data. The generator parameters are updated according to the joint loss function to optimize the quality of the synthesized data, realizing end-to-end automated training, ensuring the consistency between data realism and effect orientation, and improving generation efficiency and reliability. The optimized synthesized data is mixed with the original dataset to form an augmented dataset. The sample size is expanded through a simple and efficient mixing strategy, introducing diversity and solving the problem of poor model generalization ability caused by insufficient data. This supports the accuracy and robustness of downstream marketing models, reduces operating costs, and improves the feasibility of large-scale personalized marketing. It not only improves the intelligence and accuracy of marketing data synthesis, but also ensures long-term practicality through a closed-loop optimization mechanism, providing an efficient and reliable solution for the digital marketing field.

[0058] The above-mentioned objective three of this application is achieved through the following technical solution:

[0059] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described marketing data synthesis and enhancement method based on generative adversarial networks.

[0060] The fourth objective of this application is achieved through the following technical solution:

[0061] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described marketing data synthesis and enhancement method based on generative adversarial networks.

[0062] In summary, this application includes at least one of the following beneficial technical effects:

[0063] 1. By acquiring raw marketing datasets and extracting multimodal marketing features, this method automatically integrates multi-source data such as user profiles and behavioral sequences, solving the sparsity and heterogeneity problems of traditional marketing data, improving the comprehensiveness and accuracy of data representation, and providing high-quality input for subsequent generative adversarial networks (GANs), thereby reducing manual processing costs and strengthening the data foundation. The multimodal features are input into an effect-oriented GAN, and through a three-element structure of generator, realism discriminator, and effect discriminator, marketing objectives are directly integrated into the training process, ensuring that the synthesized data is not only realistic but also has business value, significantly improving the targeting and efficiency of marketing activities. The effect discriminator is used to evaluate the probability of the synthesized data generating positive marketing effects, generating effect evaluation signals, achieving real-time feedback and quantitative evaluation, enabling the generator to dynamically adjust its output, reducing the uncertainty of marketing activities and improving ROI. Based on the output of the realism discriminator and the effect evaluation... The algorithm calculates the joint loss function based on the estimated signal and dynamically adjusts the weight coefficients. By balancing the loss of realism and the loss of effectiveness, it adaptively optimizes model performance, avoiding overfitting or underfitting, and enhancing the robustness and scenario adaptability of the synthesized data. It updates the generator parameters based on the joint loss function, optimizing the quality of the synthesized data and achieving end-to-end automated training. This ensures consistency between data realism and effectiveness-oriented approaches, improving generation efficiency and reliability. The optimized synthesized data is then mixed with the original dataset to form an augmented dataset. This simple and efficient mixing strategy expands the sample size, introduces diversity, and solves the problem of poor model generalization caused by insufficient data. This supports the accuracy and robustness of downstream marketing models, reduces operating costs, and improves the feasibility of large-scale personalized marketing. It not only enhances the intelligence and accuracy of marketing data synthesis but also ensures long-term usability through a closed-loop optimization mechanism, providing an efficient and reliable solution for the digital marketing field.

[0064] 2. The construction of the generator, authenticity discriminator, and effect discriminator is based on a modular design, which achieves seamless integration of marketing objective vectors, making the generation process controllable and goal-driven. The effect discriminator is trained with historical data to ensure the objectivity of the evaluation. At the same time, the input combination optimizes the network training efficiency, improves the alignment between synthetic data and business needs, and enhances the practicality and scalability of the method.

[0065] 3. Dynamic weight adjustment is based on training context and data features. It optimizes loss balance in real time through heuristic rules or optimization algorithms, ensuring the best performance of the model in different scenarios and avoiding bias caused by a single objective. This improves the overall quality of synthetic data and supports the flexible implementation of complex marketing strategies.

[0066] 4. Privacy Protection: An end-to-end protection system is built through desensitization, differential noise reduction, and output inspection to ensure that the data synthesis process complies with regulatory requirements. Through quantifiable privacy budgets and adaptive adjustments, data utility and security are balanced, the risk of information leakage is reduced, the credibility and compliance of the system are enhanced, and a secure foundation is provided for the application of marketing data. Attached Figure Description

[0067] Figure 1 This is a flowchart of an implementation embodiment of the marketing data synthesis and enhancement method based on generative adversarial networks in this application;

[0068] Figure 2 This is a flowchart illustrating the implementation of step S10 in an embodiment of a marketing data synthesis and enhancement method based on generative adversarial networks.

[0069] Figure 3 This is a flowchart illustrating the implementation of step S20 in an embodiment of a marketing data synthesis and enhancement method based on generative adversarial networks.

[0070] Figure 4 This is a flowchart illustrating the implementation of step S30 in an embodiment of a marketing data synthesis and enhancement method based on generative adversarial networks.

[0071] Figure 5 This is a flowchart illustrating the implementation of step S40 in the embodiment of the marketing data synthesis and enhancement method based on generative adversarial networks in this application.

[0072] Figure 6 This is another implementation flowchart of the marketing data synthesis and enhancement method based on generative adversarial networks in this application;

[0073] Figure 7 This is a flowchart illustrating the implementation of privacy protection steps in an embodiment of the marketing data synthesis and enhancement method based on generative adversarial networks in this application.

[0074] Figure 8 This is a schematic diagram of an embodiment of the marketing data synthesis and enhancement system based on generative adversarial networks in this application;

[0075] Figure 9 This is a schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0076] The following is in conjunction with the appendix Figure 1-9 This application will be described in further detail.

[0077] In one embodiment, such as Figure 1 As shown, this application discloses a marketing data synthesis and enhancement method based on generative adversarial networks, which specifically includes the following steps:

[0078] S10: Obtain the original marketing dataset and extract multimodal marketing features based on the original marketing dataset.

[0079] In this embodiment, the original marketing dataset refers to a diverse collection of data collected from digital marketing platforms, including user profile data (such as age, gender, and interest tags), user behavior sequence data (such as clickstream and purchase history), product information data (such as product category and price), and marketing context data (such as advertising channels and timestamps). This data represents multi-dimensional information about marketing activities and is heterogeneous and dynamic. Multimodal marketing features refer to a unified feature vector extracted by fusing different modalities of data (such as text, numerical values, and sequences), which can comprehensively reflect user behavior patterns and marketing effectiveness, providing high-quality input for subsequent generative adversarial networks.

[0080] Specifically, raw marketing datasets are typically obtained from CRM systems or advertising platforms via enterprise API interfaces or database queries, and the data format may include JSON, CSV, or real-time streaming data. First, the raw data undergoes preprocessing, including data cleaning (e.g., removing duplicate records), missing value handling (using mean imputation or predictive models to fill in missing values), and outlier detection (using statistical methods such as Z-score to identify outliers), generating a standardized marketing data sequence. Then, based on the processed data sequence, multimodal features are extracted: for example, categorical features (e.g., one-hot encoding) are extracted from user profiles, and temporal features (e.g., sliding window statistics) are extracted from behavioral sequences. Attention mechanisms are then used for cross-modal alignment and fusion, mapping features from different sources to a unified vector space to generate multimodal marketing feature vectors. Simultaneously, each feature vector sample is labeled with its corresponding marketing funnel stage (e.g., awareness, consideration, conversion stage) to facilitate optimization of the generative adversarial network for specific marketing objectives. This step reduces manual intervention through automated feature engineering, improving the efficiency and consistency of data processing.

[0081] S20: Input the multimodal marketing features into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, an authenticity discriminator, and an effect discriminator.

[0082] In this embodiment, the effect-oriented generative adversarial network is an improved GAN architecture specifically designed for marketing scenarios. It adds an effect discriminator to the traditional GAN, introducing marketing objectives as the optimization guide. The generator is responsible for synthesizing new marketing data samples, the authenticity discriminator evaluates the authenticity of the data, and the effect discriminator predicts the probability of the synthesized data's effect in a specific marketing scenario.

[0083] Specifically, the Generative Adversarial Network (GAN) is constructed based on the TensorFlow or PyTorch framework: the generator input includes a random noise vector (providing diversity) and a marketing objective vector (such as target user group identifiers and expected improvement metrics), generating synthetic data through a multilayer perceptron. A realism discriminator uses a convolutional neural network to determine whether the input samples are real data. The effectiveness discriminator is a pre-trained classifier trained on historical marketing campaign data (positive samples as success cases, negative samples as failure cases), outputting the probability that the synthetic data leads to target user behavior. During training, the multimodal marketing feature vector and the marketing objective vector are combined and input into the network; adversarial training optimizes the generator parameters to ensure that the synthetic data aligns with business objectives.

[0084] S30: Use the effect discriminator to evaluate the probability that the synthetic marketing data output by the generator will produce a positive marketing effect in a predetermined marketing scenario, and generate an effect evaluation signal.

[0085] In this embodiment, the probability of positive marketing effect refers to the likelihood that synthetic data will drive target behavior (such as purchase or sharing) in a real-world scenario. The effect evaluation signal is a quantitative indicator used to guide generator optimization. The predetermined marketing scenario may include specific advertising channels or user segments to ensure the contextual relevance of the evaluation.

[0086] Specifically, after the generator outputs synthetic marketing data samples based on random noise and a marketing target vector, the effect discriminator receives these samples and calculates the conditional probability using its internal parameters (such as the weight matrix). This conditional probability is the probability that the synthetic data will lead to the target user's behavior given the marketing target vector. During the calculation, a softmax function may be used to output multi-class probabilities. This probability is then compared with a preset effect threshold (such as 0.7) to generate a binary preliminary signal (1 indicating compliance, 0 indicating non-compliance). Finally, through a predefined mapping function (such as linear scaling or a sigmoid function) combined with the target improvement magnitude (a larger magnitude results in a stronger signal), a continuous effect evaluation signal is generated. Real-time evaluation and signal generation provide a feedback mechanism to the generation process, improving the effectiveness and adaptability of the synthetic data.

[0087] S40: Based on the output of the authenticity discriminator and the effect evaluation signal, calculate the joint loss function of the generator, and dynamically adjust the weight coefficients of the authenticity loss term and the effect loss term in the joint loss function.

[0088] In this embodiment, the joint loss function is the core of generator optimization. By combining realism loss (measuring data fidelity) and effectiveness loss (measuring marketing effectiveness), a balance of multiple objectives is achieved. Dynamically adjusting the weight coefficients allows the system to adaptively prioritize realism or effectiveness based on the training stage or data sparsity, avoiding overfitting or underfitting.

[0089] Specifically, the authenticity loss term is calculated based on the output of the authenticity discriminator, such as using adversarial loss (e.g., least squares loss), to measure the proportion of generated data judged as real. The effect loss term is calculated based on the effect evaluation signal, such as mean squared error loss, to quantify the deviation of the synthetic data from the expected effect. The weight coefficients 'a' (authenticity weight) and 'β' (effect weight) are dynamically determined based on the current training epoch, the sparsity of the original dataset (e.g., increasing effect weight if the sparsity is high), or the magnitude of improvement in the marketing objective (increasing effect weight if the magnitude is large), possibly using heuristic rules or optimization algorithms (e.g., gradient descent). Then, the adjusted weights are used to calculate the total joint loss function: L_total = a * L_authenticity + β * L_effect. This dynamic weighting mechanism enhances the model's robustness and scene adaptability, ensuring a reasonable trade-off between authenticity and effect in the synthetic data.

[0090] S50: Update the parameters of the generator according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effectiveness.

[0091] In this embodiment, parameter updating is the core of the training process. The generator network weights are optimized through backpropagation, allowing the synthetic data to improve gradually. The authenticity requirement ensures that the data is not identified as fake, while the performance-oriented requirement guarantees that the data can drive business metrics. The combination of these two aspects enhances the practical value and reliability of the synthetic data.

[0092] Specifically, an optimizer (such as Adam or SGD) is used to calculate gradients based on the joint loss function and update the generator's parameters (such as weights and biases). The training process is iterative, with the generator producing a batch of synthetic data in each round, calculating the loss, and then backpropagating to adjust the parameters. After optimization, the synthetic marketing data approximates real data in terms of visual and statistical characteristics. Simultaneously, it is validated by an effectiveness discriminator to ensure positive effects in the intended marketing scenario. Through end-to-end optimization, the data generation process is automated and intelligent, reducing the need for manual parameter tuning.

[0093] S60: The optimized synthetic marketing data is mixed with the original marketing dataset to form an enhanced marketing dataset.

[0094] In this embodiment, data augmentation aims to expand the scale of training samples to address the problem of insufficient data, while introducing diversity through synthetic data to improve the generalization ability of downstream models (such as recommendation systems or prediction models).

[0095] Specifically, the blending operation is implemented through a data pipeline: first, the synthetic data is sampled to ensure its proportion is balanced with the original data (e.g., synthetic data accounts for 20%), and then it is concatenated or interpolated with the original dataset. The enhanced dataset is used to train the marketing model.

[0096] In this embodiment, by acquiring the original marketing dataset and extracting multimodal marketing features, this method automatically integrates multi-source data such as user profiles and behavioral sequences, solving the sparsity and heterogeneity problems of traditional marketing data, improving the comprehensiveness and accuracy of data representation, and providing high-quality input for subsequent generative adversarial networks, thereby reducing manual processing costs and strengthening the data foundation. The multimodal features are input into an effect-oriented generative adversarial network. Through a three-element structure of generator, realism discriminator, and effect discriminator, marketing objectives are directly integrated into the training process, ensuring that the synthesized data is not only realistic but also has business value, significantly improving the targeting and efficiency of marketing activities. The effect discriminator is used to evaluate the probability of the synthesized data generating positive marketing effects, generating effect evaluation signals, achieving real-time feedback and quantitative evaluation, enabling the generator to dynamically adjust its output, reducing the uncertainty of marketing activities and improving the return on investment. Based on the output of the realism discriminator and the effect... The algorithm calculates a joint loss function based on the evaluation signal and dynamically adjusts the weight coefficients. By balancing the loss of realism and the loss of effectiveness, it adaptively optimizes model performance, avoiding overfitting or underfitting, and enhancing the robustness and scenario adaptability of the synthesized data. The generator parameters are updated based on the joint loss function to optimize the quality of the synthesized data, achieving end-to-end automated training. This ensures consistency between data realism and effectiveness-oriented approaches, improving generation efficiency and reliability. The optimized synthesized data is then mixed with the original dataset to form an augmented dataset. This simple and efficient mixing strategy expands the sample size, introduces diversity, and solves the problem of poor model generalization caused by insufficient data. This supports the accuracy and robustness of downstream marketing models, reduces operating costs, and improves the feasibility of large-scale personalized marketing. It not only enhances the intelligence and accuracy of marketing data synthesis but also ensures long-term usability through a closed-loop optimization mechanism, providing an efficient and reliable solution for the digital marketing field.

[0097] In one embodiment, such as Figure 2 As shown, in step S10, the original marketing dataset is obtained, and multimodal marketing features are extracted based on the original marketing dataset, specifically including:

[0098] S11: Obtain a raw marketing dataset containing at least two of the following: user profile data, user behavior sequence data, product information data, and marketing context data.

[0099] In this embodiment, user profile data refers to structured information describing user attributes, such as demographic characteristics and preference tags; user behavior sequence data refers to the interaction records between users and marketing content, arranged in chronological order; product information data includes product details and pricing; and marketing context data captures environmental factors for ad placement.

[0100] Specifically, data is acquired in real-time or in batches from multiple sources (such as website analytics tools and social media platforms) using data integration tools (such as Apache Kafka or custom APIs). Format validation is performed during data acquisition to ensure compatibility; for example, image description text is converted into numerical features.

[0101] S12: Perform data cleaning, missing value processing, and outlier detection on the original marketing dataset to obtain a standardized marketing data sequence.

[0102] In this embodiment, data cleaning aims to remove noise and inconsistent data, missing value handling prevents model bias, and outlier detection maintains data quality.

[0103] Specifically, Python libraries such as Pandas are used for data cleaning: duplicates are removed, missing values ​​are filled, and outliers are detected using the Isolation Forest algorithm. The output is a standardized data sequence, such as time series normalized to the 0-1 range, ensuring data consistency and model stability.

[0104] S13: Based on the processed marketing data sequence, extract multimodal marketing data features, use an attention mechanism to perform cross-modal alignment and fusion of the extracted multimodal marketing data features, generate a unified multimodal marketing feature vector, and label each feature vector sample with its corresponding marketing funnel stage label.

[0105] In this embodiment, the attention mechanism is a deep learning technique that calculates the weights of different modal features to achieve focused attention and automatic alignment, avoiding information redundancy. Marketing funnel stage labels are defined based on business rules; for example, "awareness stage" corresponds to low-interaction users, and "conversion stage" corresponds to high-value users.

[0106] Specifically, feature extraction uses neural network models: for example, convolutional neural networks process image-related contextual data, and recurrent neural networks process behavioral sequences. Attention mechanisms calculate inter-modal similarity, and weighted summations generate a fused vector. Labeling is automatically performed using a rule engine or clustering algorithm, improving feature interpretability and model-oriented design.

[0107] In one embodiment, such as Figure 3 As shown, in step S20, namely the effect-oriented generative adversarial network, its construction and input specifically include:

[0108] S21: Construct the generator, whose input includes a random noise vector and a marketing target vector, the marketing target vector being used to characterize the desired marketing objective, including the target user group identifier, the type of marketing metric to be improved, and the target improvement amount.

[0109] In this embodiment, a random noise vector introduces randomness to enhance data diversity, while a marketing objective vector quantifies business needs, making the generation process controllable. The target improvement margin, for example, is a specified 10% increase in conversion rate, ensuring the synthesized data has a clear direction for optimization.

[0110] Specifically, the generator employs a generative model such as GAN or VAE. The input layer receives a concatenation of noise and the target vector, and synthesizes features through a fully connected layer. The marketing target vector is dynamically generated by the business system and embedded into the network to guide data synthesis.

[0111] S22: Construct the authenticity discriminator to receive real marketing data samples or synthetic marketing data samples generated by the generator, and output the probability that the sample is a real sample.

[0112] Specifically, the realism discriminator, as the core of adversarial training, improves the realism of the generated data through binary classification tasks and avoids generating obviously fake samples. The discriminator uses a discriminative model such as CNN, and adopts the cross-entropy loss function during training to evaluate the realism of the samples in real time and feed it back to the generator for iterative optimization.

[0113] S23: Construct the effect discriminator, which is trained based on historical marketing campaign data, including positive and negative samples. The effect discriminator is used to evaluate the predictive probability that the input data sample can lead to the occurrence of target user behavior in a given marketing scenario.

[0114] In this embodiment, the effect discriminator incorporates domain knowledge to quantify marketing effectiveness as probability values. Positive samples may come from high-conversion activities, while negative samples come from inefficient activities, ensuring the objectivity of the evaluation.

[0115] Specifically, the effect discriminator is trained using logistic regression or a neural network. After inputting synthetic data, it outputs probability values ​​as part of the generator's loss function, directly optimizing marketing effectiveness.

[0116] S24: The multimodal marketing feature vector is combined with the corresponding marketing target vector and then input into the generative adversarial network for training.

[0117] Specifically, the combined input ensures that the network processes data features and business objectives simultaneously, achieving end-to-end optimization. The training process uses mini-batch gradient descent, gradually improving the quality of synthetic data by alternately optimizing the generator and discriminator. The number of training rounds is dynamically adjusted according to the data scale.

[0118] In one embodiment, such as Figure 4 As shown, in step S30, the probability that the synthetic marketing data output by the generator will produce a positive marketing effect under the predetermined marketing scenario is evaluated using an effect discriminator, and an effect evaluation signal is generated. Specifically, this includes:

[0119] S31: Input the synthetic marketing data sample generated by the generator based on the random noise vector and the marketing target vector into the effect discriminator.

[0120] Specifically, sample input is implemented through memory or distributed caching, ensuring low latency and high throughput, realizing pipelined operations for generation and evaluation, and ensuring seamless data transfer.

[0121] S32: The effect discriminator calculates the conditional probability of the synthetic marketing data sample corresponding to the target user behavior specified in the marketing target vector based on its internal model parameters, as the probability of the positive marketing effect.

[0122] In this embodiment, the conditional probability reflects the strength of the association between the data and the target, and the internal parameters are obtained by training with historical data to ensure prediction accuracy.

[0123] Specifically, the calculation uses a probabilistic model such as a Bayesian network, and the output probability value is used as the basis for optimization.

[0124] S33: Compare the calculated conditional probability with the preset effect threshold to generate a binarized preliminary effect evaluation signal.

[0125] Specifically, the threshold setting is based on business risk preference, binarization simplifies the decision-making process, comparison operations are implemented through if-else logic, and the generated signal is used for loss calculation.

[0126] S34: Based on the continuous value of the conditional probability or the binarized preliminary effect evaluation signal, and combined with the target improvement magnitude in the marketing target vector, the final effect evaluation signal is generated through a predefined mapping function.

[0127] Specifically, the mapping function defines the signal scaling rules, adjusts the signal strength according to the target increase, and achieves dynamic optimization. The mapping function can be a piecewise linear function to ensure that the signal is consistent with the business objective.

[0128] In one embodiment, such as Figure 5 As shown, in step S40, based on the output of the realism discriminator and the effect evaluation signal, the joint loss function of the generator is calculated, and the weight coefficients of the realism loss term and the effect loss term in the joint loss function are dynamically adjusted, specifically including:

[0129] S41: Calculate the authenticity loss term of the generator relative to the authenticity discriminator, the authenticity loss term being calculated based on the discrimination result of the authenticity discriminator on the synthetic marketing data.

[0130] In this embodiment, the authenticity loss term drives the generated data to approximate the true distribution, which is the basis for GAN training.

[0131] Specifically, the loss calculation uses standard GAN loss functions, such as binary cross-entropy, to optimize the generator's ability to deceive the discriminator.

[0132] S42: Calculate the effect loss term of the generator relative to the effect discriminator, the effect loss term being calculated based on the effect evaluation signal.

[0133] Specifically, the effect loss item incorporates business objectives into the optimization, directly improving the usability of the synthesized data. The loss calculation can use L1 or L2 loss to minimize the gap with the target effect.

[0134] S43: Based on the current training stage, the sparsity of the original marketing dataset, or the target improvement specified in the marketing target vector, dynamically determine the weight coefficients α and β of the authenticity loss term and the effect loss term in the joint loss function.

[0135] In this embodiment, dynamic adjustment enables self-learning capabilities, such as prioritizing realism in early training and prioritizing effectiveness in later training.

[0136] Specifically, the weights are automatically adjusted based on monitoring metrics (such as the loss curve).

[0137] S44: Calculate the total joint loss function of the generator using the dynamically adjusted weighting coefficients α and β.

[0138] In one embodiment, such as Figure 6 As shown, after step S60, that is, after mixing the optimized synthetic marketing data with the original marketing dataset to form the enhanced marketing dataset, the marketing data synthesis and enhancement method based on generative adversarial networks further includes:

[0139] S70: Obtain marketing performance feedback data based on the enhanced marketing dataset.

[0140] Specifically, feedback data comes from real marketing campaigns, such as click-through rates or conversion rates, to verify the validity of synthetic data. Feedback is collected through monitoring tools to form a closed-loop optimization foundation.

[0141] S80: The marketing performance feedback data is fed back to the performance discriminator to update the parameters of the performance discriminator, forming a closed-loop optimization system.

[0142] Specifically, the feedback process, through data stream processing, regularly updates the performance discriminator. The closed-loop system enables continuous learning, allowing the model to adapt to changing market environments and improve long-term performance.

[0143] In one embodiment, such as Figure 7 As shown, the marketing data synthesis and enhancement method based on generative adversarial networks also includes privacy protection steps:

[0144] S101: Before inputting raw marketing data into the feature extraction process, delete or desensitize direct identifiers.

[0145] In this embodiment, a direct identifier refers to an information field that can individually identify a specific individual, such as a user's name, ID number, mobile phone number, or email address. Deletion or anonymization aims to eliminate the risk of direct exposure of personal identity at the source, ensuring that subsequent data processing complies with the basic requirements of privacy regulations (such as GDPR and CCPA).

[0146] Specifically, the processing is implemented through automated scripts or a data management platform: First, the system scans the metadata structure of the original marketing data to identify predefined direct identifier fields (such as field names containing keywords like "name" or "ID"). Then, based on the privacy policy configuration, the system performs deletion or desensitization operations on the identifiers: deletion directly removes the field and its data; desensitization uses a hash function (such as SHA-256) for irreversible encryption, or uses masking techniques (such as displaying only the last four digits of the phone number). For example, a user's email address "user@example.com" becomes a fixed-length string "5e884898da..." after hashing, which retains some of the data's utility (such as for deduplication) while avoiding direct identification. Only after processing does the data enter the feature extraction process, ensuring that subsequent generative adversarial network training does not access the original identifier information.

[0147] S102: During the training process of the generator, noise that meets the differential privacy requirements is added to the random noise vector input to it, and the scale of the noise is adaptively adjusted according to a preset privacy budget.

[0148] In this embodiment, differential privacy is a rigorous mathematical privacy framework that adds controllable noise to the data or model to ensure that the addition or removal of a single data point does not significantly affect the output, thereby preventing the inference of original individual information from synthetic data. Random noise vectors serve as the generator's input source; adding noise blurs the connection between the generation process and specific training data. A privacy budget quantifies the strength of privacy protection; a smaller budget provides stronger protection but may reduce data utility. An adaptive adjustment mechanism dynamically optimizes the noise scale based on the training stage and sensitivity, balancing privacy and data quality.

[0149] Specifically, the implementation process is integrated into the training loop of the generative adversarial network (GAN): each time the generator receives a random noise vector, the system generates a noise vector conforming to a Laplace or Gaussian distribution (with a mean of zero, and a scale determined by the privacy budget ε). The noise scale is calculated using a formula (e.g., scale = sensitivity / ε), where sensitivity depends on the maximum possible impact on the training data. The privacy budget ε is preset by the administrator (e.g., ε = 1.0) and can be adaptively adjusted based on data sensitivity (e.g., a smaller ε for medical marketing data). For example, in the early stages of training, the system may use a larger ε (e.g., 2.0) to add less noise to accelerate convergence; as training progresses, ε is gradually decreased (e.g., down to 0.5) to increase the noise intensity and strengthen privacy protection. The vector with added noise is then input into the generator, ensuring that the synthesized data does not remember the original individual records, while providing quantifiable privacy protection through the mathematical guarantees of differential privacy.

[0150] S103: The synthesized data output by the generator is checked, and the synthesized marketing data is only output for subsequent enhancement and application steps after it meets privacy and security standards.

[0151] Specifically, privacy and security checks are the final checkpoint for data output, ensuring through multi-dimensional verification that the synthetic data does not contain privacy remnants or inferred risks. Privacy and security standards include, but are not limited to: a similarity threshold between the synthetic data and the original training data (to prevent overfitting of individuals), re-identification risk assessment (e.g., testing through attack models), and compliance indicators (e.g., whether it conforms to industry standards). After the generator outputs synthetic marketing data, the audit module first calculates its statistical distance from the original dataset (e.g., using Jensen-Shannon divergence) to ensure the similarity is below a preset threshold (e.g., 0.1), preventing the synthetic data from approximately replicating the original individuals. Then, a re-identification attack simulation is performed: an attacker model attempts to match the synthetic data to the original identifiers. If the matching success rate exceeds a safety threshold (e.g., <5%), the data is deemed high-risk. Simultaneously, the data format and content are checked to ensure compliance with privacy policies (e.g., whether there are unintentional de-identification remnants). Only after all checks pass are the synthetic data marked as safe and allowed to be mixed with the original dataset for augmentation; otherwise, the data is discarded or returned to the generator for re-optimization.

[0152] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0153] In one embodiment, a marketing data synthesis and enhancement system based on generative adversarial networks (GANs) is provided. This system corresponds one-to-one with the marketing data synthesis and enhancement methods based on GANs described in the above embodiments. Figure 8 As shown, this marketing data synthesis and enhancement system based on generative adversarial networks includes:

[0154] Marketing data processing module: used to acquire the original marketing dataset and extract multimodal marketing features based on the original marketing dataset;

[0155] Generative Adversarial Network Training Module: Used to input the multimodal marketing features into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, a realism discriminator, and an effect discriminator;

[0156] Marketing effectiveness evaluation module: used to evaluate the probability that the synthetic marketing data output by the generator will produce positive marketing effects in a predetermined marketing scenario using the effectiveness discriminator, and generate an effectiveness evaluation signal;

[0157] Dynamic adjustment module: used to calculate the joint loss function of the generator based on the output of the realism discriminator and the effect evaluation signal, and dynamically adjust the weight coefficients of the realism loss term and the effect loss term in the joint loss function;

[0158] The data optimization module is used to update the parameters of the generator according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effect orientation.

[0159] The marketing data enhancement module is used to mix the optimized synthetic marketing data with the original marketing dataset to form an enhanced marketing dataset.

[0160] Specific limitations regarding the marketing data synthesis and enhancement system based on generative adversarial networks (GANs) can be found in the limitations of the marketing data synthesis and enhancement method based on GANs mentioned above, and will not be repeated here. Each module in the aforementioned marketing data synthesis and enhancement system based on GANs can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0161] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and the database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores marketing data and generative adversarial networks (GANs). The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a marketing data synthesis and enhancement method based on a GAN.

[0162] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:

[0163] Obtain the original marketing dataset and extract multimodal marketing features based on the original marketing dataset;

[0164] The multimodal marketing features are input into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, an authenticity discriminator, and an effect discriminator;

[0165] The effect discriminator is used to evaluate the probability that the synthetic marketing data output by the generator will produce a positive marketing effect in a predetermined marketing scenario, and an effect evaluation signal is generated.

[0166] Based on the output of the authenticity discriminator and the effect evaluation signal, the joint loss function of the generator is calculated, and the weight coefficients of the authenticity loss term and the effect loss term in the joint loss function are dynamically adjusted.

[0167] The generator parameters are updated according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effectiveness.

[0168] The optimized synthetic marketing data is then mixed with the original marketing dataset to form an enhanced marketing dataset.

[0169] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0170] Obtain the original marketing dataset and extract multimodal marketing features based on the original marketing dataset;

[0171] The multimodal marketing features are input into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least a generator, an authenticity discriminator, and an effect discriminator;

[0172] The effect discriminator is used to evaluate the probability that the synthetic marketing data output by the generator will produce a positive marketing effect in a predetermined marketing scenario, and an effect evaluation signal is generated.

[0173] Based on the output of the authenticity discriminator and the effect evaluation signal, the joint loss function of the generator is calculated, and the weight coefficients of the authenticity loss term and the effect loss term in the joint loss function are dynamically adjusted.

[0174] The generator parameters are updated according to the joint loss function to optimize the quality of the synthetic marketing data, so that the optimized synthetic marketing data simultaneously meets the requirements of authenticity and marketing effectiveness.

[0175] The optimized synthetic marketing data is then mixed with the original marketing dataset to form an enhanced marketing dataset.

[0176] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0178] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A marketing data synthesis and augmentation method based on a generative adversarial network, characterized in that, The marketing data synthesis and enhancement method based on the generative adversarial network comprises the steps of: obtaining an original marketing data set, extracting multi-modal marketing features based on the original marketing data set; inputting the multi-modal marketing features into an effect-oriented generative adversarial network, wherein the generative adversarial network comprises at least one generator, one authenticity discriminator and one effect discriminator; evaluating the probability of the synthesized marketing data output by the generator in a predetermined marketing scenario to produce a positive marketing effect by using the effect discriminator, and generating an effect evaluation signal; based on the output of the authenticity discriminator and the effect evaluation signal, calculating the joint loss function of the generator, and dynamically adjusting the weight coefficients of the authenticity loss term and the effect loss term in the joint loss function, specifically including: calculating the authenticity loss term of the generator relative to the authenticity discriminator, which is calculated based on the discrimination result of the authenticity discriminator on the synthesized marketing data; calculating the effect loss term of the generator relative to the effect discriminator, which is calculated based on the effect evaluation signal; dynamically determining the weight coefficients α and β of the authenticity loss term and the effect loss term in the joint loss function according to the current training stage, the data sparsity degree of the original marketing data set or the target promotion amplitude specified in the marketing target vector; using the dynamically adjusted weight coefficients α and β to calculate the total joint loss function of the generator; updating the parameters of the generator according to the joint loss function to optimize the quality of the synthesized marketing data, so that the optimized synthesized marketing data meets the authenticity requirement and the marketing effect-oriented requirement at the same time; mixing the optimized synthesized marketing data with the original marketing data set to form an enhanced marketing data set. 2.The marketing data synthesis and enhancement method based on a generative adversarial network according to claim 1, characterized in that: The original marketing data set is obtained, and multi-modal marketing features are extracted based on the original marketing data set, specifically including: obtaining an original marketing data set containing at least two of user portrait data, user behavior sequence data, product information data and marketing context data; performing data cleaning, missing value processing and outlier detection on the original marketing data set to obtain standardized marketing data sequences; based on the processed marketing data sequences, multi-modal marketing data features are extracted, the extracted multi-modal marketing data features are aligned and fused across modalities by using an attention mechanism, a unified multi-modal marketing feature vector is generated, and each feature vector sample is labeled with its corresponding marketing funnel stage label. 3.The marketing data synthesis and enhancement method based on a generative adversarial network according to claim 2, characterized in that: The effect-oriented generative adversarial network comprises: constructing the generator, which inputs include random noise vector and marketing target vector, the marketing target vector is used to represent the expected marketing target, including target user group identification, expected promotion marketing index type and target promotion amplitude; constructing the authenticity discriminator, which is used to receive real marketing data samples or synthesized marketing data samples generated by the generator, and output the probability that the sample is a real sample; The effect discriminator is constructed based on historical marketing activity data containing positive and negative samples, and is used to evaluate the prediction probability of input data samples in a given marketing scenario to cause target user behavior to occur; After the multi-modal marketing feature vector and the corresponding marketing target vector are combined, they are input into the generative adversarial network for training.

4. The marketing data synthesis and enhancement method based on a generative adversarial network according to claim 3, characterized in that: The effect discriminator evaluates the probability of the synthetic marketing data output by the generator to produce a positive marketing effect in a predetermined marketing scenario, and generates an effect evaluation signal, specifically including: The synthetic marketing data sample generated by the generator based on the random noise vector and the marketing target vector is input into the effect discriminator; The effect discriminator calculates the conditional probability of the synthetic marketing data sample corresponding to the target user behavior specified in the marketing target vector based on its internal model parameters as the probability of the positive marketing effect; The calculated conditional probability is compared with a preset effect threshold to generate a binary preliminary effect evaluation signal; Based on the continuous value of the conditional probability or the binary preliminary effect evaluation signal, the target improvement amplitude in the marketing target vector is combined to generate the final effect evaluation signal through a predefined mapping function. 5.The marketing data synthesis and enhancement method based on a generative adversarial network according to claim 1, wherein: After the optimized synthetic marketing data is mixed with the original marketing data set to form an enhanced marketing data set, it further includes: Based on the enhanced marketing data set, marketing effect feedback data is obtained; The marketing effect feedback data is fed back to the effect discriminator for updating the parameters of the effect discriminator to form a closed-loop optimization system. 6.The marketing data synthesis and enhancement method based on a generative adversarial network according to claim 1, wherein: The marketing data synthesis and enhancement method based on the generative adversarial network further includes: Before the original marketing data is input into the feature extraction process, the direct identifier is deleted or desensitized; In the training process of the generator, noise meeting the differential privacy requirement is added to the random noise vector input thereto, and the scale of the noise is adaptively adjusted according to a preset privacy budget; The synthetic data output by the generator is checked, and only when the privacy security standard is met, the synthetic marketing data is output for subsequent enhancement and application steps.

7. A marketing data synthesis and augmentation system based on a generative adversarial network, characterized in that, It includes: A marketing data processing module is used to obtain an original marketing data set and extract multi-modal marketing features based on the original marketing data set; A generative adversarial network training module is used to input the multi-modal marketing features into an effect-oriented generative adversarial network, wherein the generative adversarial network includes at least one generator, one reality discriminator and one effect discriminator; A marketing effect evaluation module is used to evaluate the probability of the synthetic marketing data output by the generator to produce a positive marketing effect in a predetermined marketing scenario using the effect discriminator, and generate an effect evaluation signal. The dynamic adjustment module is configured to calculate a joint loss function of the generator based on the output of the authenticity discriminator and the effect evaluation signal, and dynamically adjust weight coefficients of a authenticity loss term and an effect loss term in the joint loss function. Specifically, the dynamic adjustment module is configured to calculate a authenticity loss term of the generator with respect to the authenticity discriminator, the authenticity loss term being calculated based on a discrimination result of the authenticity discriminator on the synthetic marketing data; calculate an effect loss term of the generator with respect to the effect discriminator, the effect loss term being calculated based on the effect evaluation signal; dynamically determine weight coefficients a and β of the authenticity loss term and the effect loss term in the joint loss function according to a current training stage, a data sparsity degree of the original marketing data set, or a specified target promotion amplitude in a marketing target vector; and calculate a total joint loss function of the generator using the dynamically adjusted weight coefficients a and β. The data optimization module is configured to update parameters of the generator according to the joint loss function, so as to optimize a quality of the synthetic marketing data, so that the optimized synthetic marketing data meets both authenticity requirements and marketing effect-oriented requirements. The marketing data enhancement module is configured to mix the optimized synthetic marketing data with the original marketing data set to form an enhanced marketing data set.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the marketing data synthesis and enhancement method based on the generative adversarial network according to any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to implement the steps of the marketing data synthesis and enhancement method based on the generative adversarial network according to any one of claims 1 to 6.

Images