Generative adversarial network-based e-commerce platform marketing strategy generation method

Abstract

The invention discloses an e-commerce platform marketing strategy generation method based on a generative adversarial network. The method comprises the following steps: S1, collecting user behavior data to construct a dynamic behavior data sequence; s2, performing semi-supervised hierarchical clustering on the dynamic behavior data sequence by adopting an R3GAN model; s3, constructing a staged progressive R3GAN strategy generation model according to the clustering group; s4, constructing a double-layer flying fox optimization algorithm to optimize strategy parameters; s5, deploying a marketing strategy and monitoring user response in real time; s6, constructing a strategy self-sensing adaptation mechanism, and dynamically adjusting a model parameter space; and S7, setting a strategy performance threshold value, automatically decommissioning and redeploying. According to the invention, the dynamic adaptability and accuracy of the marketing strategy are improved.

Description

Technical Field

[0001] The present invention relates to the field of e-commerce intelligent marketing technology, and in particular to a method for generating marketing strategies for e-commerce platforms based on a generative adversarial network. Background Art

[0002] With the rapid development of internet technology, e-commerce platforms have become a vital channel for product marketing. In recent years, e-commerce platforms have gradually incorporated artificial intelligence and machine learning technologies, analyzing historical user behavior data to achieve precise and personalized marketing strategies. Generative adversarial networks (GANs), in particular, have been widely researched and applied in marketing content generation due to their ability to generate diverse data. Currently, some e-commerce platforms have begun utilizing GAN-related technologies to mine and analyze historical interaction data such as user browsing behavior, purchasing behavior, and click patterns, thereby generating corresponding marketing recommendation strategies to improve user satisfaction and marketing effectiveness.

[0003] However, existing e-commerce platform marketing strategy generation methods typically use single-stage, static data analysis strategies, lacking an agile response mechanism to real-time dynamic changes in user behavior. This makes it difficult to quickly capture changing trends in user preferences, resulting in insufficient personalization of generated marketing strategies and an inability to accurately and effectively adapt to users' real-time needs. Furthermore, traditional generative adversarial network models commonly suffer from problems such as pattern collapse and strategy homogeneity during strategy generation, further limiting the innovation and diversity of marketing strategies. Furthermore, existing technologies often use conventional optimization methods during strategy parameter optimization, resulting in slow parameter adjustment and an inability to effectively respond to rapid changes in the actual marketing environment. This results in slow strategy response and continued performance degradation.

[0004] Therefore, how to provide a method for generating marketing strategies for e-commerce platforms based on generative adversarial networks is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0005] One purpose of the present invention is to propose a method for generating marketing strategies for e-commerce platforms based on generative adversarial networks. The present invention has the effect of responding to changes in user needs in real time and improving the personalization and accuracy of marketing strategies.

[0006] According to an embodiment of the present invention, a method for generating marketing strategies for an e-commerce platform based on a generative adversarial network includes the following steps:

[0007] S1. Collect user behavior data on the e-commerce platform, pre-process the behavior data, and construct a user dynamic behavior data sequence based on time series;

[0008] S2. Use the R3GAN model to perform semi-supervised hierarchical clustering on the user dynamic behavior data sequence to form a hierarchical cluster group with features that change dynamically over time;

[0009] S3. Based on the hierarchical clustering of features, a phased and progressive R3GAN strategy generation model is constructed to generate coarse-grained marketing strategies and fine-grained marketing strategies respectively;

[0010] S4. Parameterize the fine-grained marketing strategy, construct a double-layer nested Feihu optimization algorithm, and optimize the fine-grained marketing strategy;

[0011] S5. Deploy the optimized fine-grained marketing strategy in the e-commerce platform operating environment, monitor actual user response data in real time, and record the difference between the user's real-time response and the expected response;

[0012] S6. Through the strategy self-perception adaptation mechanism, the parameter space of the phased progressive R3GAN strategy generation model is actively adjusted in real time based on the difference data, so that the next round of marketing strategy generation process can autonomously adapt to the real-time user response changes;

[0013] S7. Set a performance attenuation threshold for the marketing strategy and monitor the performance of the marketing strategy in real time. When the sustained performance falls below the attenuation threshold, the retirement mechanism of the marketing strategy is automatically triggered, and the process of steps S3 to S6 is immediately restarted to generate and deploy a new replacement marketing strategy.

[0014] Optionally, the S1 specifically includes:

[0015] S11. During each page visit in the user's browsing history, the page scroll position changes and mouse movement trajectory are captured in real time, and combined with the page dwell time, page focus point trajectory data is established;

[0016] S12. Identify and extract the user's search intent change process based on the user's search keywords and the secondary modified search behavior after each search, and construct search intent evolution data reflecting the continuous evolution of search intent;

[0017] S13. For each click event in the number of user clicks, extract the positional relationship between the click and the page element and the page element attributes to obtain click spatial feature data of the click action;

[0018] S14. Extracting dynamic hesitation behavior data of the user's shopping decision based on the time interval between adding, removing, modifying the quantity of the shopping cart, and final purchase completion in the user's purchase behavior to form purchase hesitation behavior data;

[0019] S15. Based on the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data, identify and filter out abnormal patterns in temporal features, retaining the normal fluctuation characteristics of dynamic behavior as the user behavior dynamic sequence;

[0020] S16. Using the historical variation ranges of the user behavior dynamic sequences, respectively set the normalized upper and lower limits of dynamic adjustment to construct a user dynamic behavior data sequence that reflects the dynamic trends of the user's real behavior.

[0021] Optionally, the S2 specifically includes:

[0022] S21. Align the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data in the user dynamic behavior data sequence according to timestamps to form a time-synchronized dynamic behavior feature vector.

[0023] S22. Randomly select a preset proportion of dynamic behavior feature vectors that have been labeled with user interest categories as the supervision sample set in the clustering process, and the dynamic behavior feature vectors that have not been labeled with user interest categories as the sample set to be classified;

[0024] S23. Using the generator in the R3GAN model, with the dynamic behavior feature vectors and corresponding interest category labels of the supervised sample set as the benchmark, multiple parallel generation channels are constructed. Each channel introduces Gaussian random noise of different variances through the feature noise input layer, and generates interest category prediction labels for the sample set to be classified channel by channel. The label difference distribution between channels is used to dynamically adjust the output weights of each channel in the generator, forming a preliminary category prediction result with a difference enhancement effect.

[0025] S24. Using the discriminator in the R3GAN model, the dynamic behavior feature vectors of the supervised sample set that have been labeled with the interest category label are used as positive samples, and the dynamic behavior feature vectors of the sample set to be classified corresponding to the preliminary category prediction results are used as negative samples. The local feature differences and global feature differences of the corresponding categories of the positive and negative samples are calculated respectively, and the sum of the weighted local feature differences and the global feature differences is used as the category matching score. Based on the category matching score, the category boundary is dynamically corrected through the adaptive threshold iteration method, and the final category matching score is output.

[0026] S25. Determine the final category label of the sample set to be classified based on the final category matching score, and complete the preliminary clustering of the dynamic behavior feature vector in combination with the supervision sample set;

[0027] S26. Based on the continuous change characteristics of the dynamic behavior feature vector on the time axis, the categories of the preliminary clustering are dynamically stratified to form feature-stratified cluster groups that change dynamically over time.

[0028] Optionally, the S3 specifically includes:

[0029] S31, setting a granularity classification standard for the feature hierarchical cluster group according to the number of samples and the degree of dynamic behavior difference of different feature categories in the feature hierarchical cluster group, and dividing the cluster group into a coarse-grained category group and a fine-grained category group;

[0030] S32. Using the dynamic behavior feature vectors within the coarse-grained category groups as training samples, the generator network of the first-stage R3GAN strategy generation model is constructed. By introducing multiple groups of high-dimensional random noise vectors with different variances into the input layer of the dynamic behavior feature vectors, the initial coarse-grained marketing strategy candidate set is generated based on the dynamic combination of the transformation factors of each random noise.

[0031] S33, using the discriminator network of the first-stage R3GAN strategy generation model to determine a coarse-grained marketing strategy that adapts to the initial user demand trend;

[0032] S34. Subdivide the coarse-grained marketing strategy into a set of strategy feature sub-vectors through nonlinear mapping of the strategy feature dimension, which serves as the input benchmark of the second-stage R3GAN strategy generation model, and uses the dynamic behavior feature vectors within the fine-grained category group as supervision samples;

[0033] S35. In the generator network of the second-stage R3GAN strategy generation model, local perturbation random noise with gradually decreasing scale is introduced on each feature dimension of the strategy feature sub-vector set, gradually generating a number of fine-grained marketing strategy candidate sets with differentiated feature details;

[0034] S36. Through the discriminator network of the second-stage R3GAN strategy generation model, the real user dynamic behavior feature vectors in the fine-grained category group are used as positive samples, and the generated fine-grained marketing strategy candidate set is used as negative samples. The local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix between the positive and negative samples are calculated respectively, and the Frobenius norm of the two difference distribution matrices is used as the scoring indicator of the strategy category matching degree. Then, feedback iteration is performed based on the score gradient descent method to finally determine the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

[0035] Optionally, the first-stage R3GAN strategy generation model specifically includes a feature adaptive enhancement input module, a random noise multi-scale interaction generation module, a strategy time-space collaborative fusion module, a dynamic strategy self-confrontation module, and an adaptive momentum feedback iteration module:

[0036] The feature adaptive enhancement input module is used to receive the dynamic behavior feature vector within the coarse-grained category group, and based on the internal distance difference of the feature vector space, adaptively amplify the feature value of the key feature dimension within the dynamic behavior feature vector through the feature enhancement factor;

[0037] The random noise multi-scale interactive generation module uses multi-scale Gaussian random noise and generates multiple groups of scale interactive random noise vectors through an inter-scale interactive fusion process;

[0038] The strategic temporal-spatial collaborative fusion module inputs the enhanced features of the feature adaptive enhancement input module and the scale-interactive random noise vector generated by the random noise multi-scale interaction generation module, and forms a collaborative fusion process of strategic features by respectively constructing a temporal attention mechanism and a spatial attention mechanism, including generating temporal attention weights through a bidirectional gated recurrent unit network and generating spatial attention weights through a two-dimensional convolutional network, and then gradually fusing the temporal attention weights and the spatial attention weights to obtain a fused feature expression;

[0039] The dynamic strategy self-adversarial module generates an initial coarse-grained marketing strategy candidate set based on the fusion feature expression, randomly groups the strategies in the coarse-grained marketing strategy candidate set, performs self-adversarial training on the strategy features between each two, and dynamically adjusts the generator network parameters through the adversarial loss function between the alternative strategies;

[0040] The adaptive momentum feedback iteration module calculates the matching scores between the strategies in the marketing strategy alternative set and the real features based on the real user dynamic behavior feature vector, adaptively adjusts the momentum parameters according to the score gradient change, dynamically updates the generator network parameters, and outputs a coarse-grained marketing strategy that ultimately adapts to the initial user demand trend.

[0041] Optionally, the second-stage R3GAN policy generation model specifically includes a policy feature fine-grained recoding module, a policy dimension step-by-step perturbation module, a multi-head adaptive attention fusion module, a cross-granularity policy feature adversarial module, and a dynamic threshold feedback iteration module:

[0042] The strategy feature fine-grained recoding module is used to receive a set of strategy feature sub-vectors and, based on a feature difference matrix of dynamic behavior feature vectors within a fine-grained category group, perform dimension-by-dimension adaptive fine-grained recoding on the strategy feature sub-vectors through feature recoding to obtain fine-grained strategy initial coding features;

[0043] The strategy dimension step-by-step perturbation module adopts Gaussian random noise with decreasing scale, and embeds the fine-grained strategy initial coding features dimension by dimension through the dimension perturbation factor to obtain the dimension perturbation feature;

[0044] The multi-head adaptive attention fusion module receives dimensional perturbation features, extracts strategy temporal attention weights, dimensional feature importance weights, and feature interaction weights by constructing a multi-head adaptive attention network structure, and dynamically fuses them according to the weight vector set to obtain a multi-head fusion feature expression;

[0045] The cross-granularity strategy feature adversarial module generates a fine-grained marketing strategy candidate set from multi-head fusion feature expressions, and performs cross-granularity feature adversarial training with the coarse-grained marketing strategies output by the first-stage R3GAN strategy generation model, dynamically adjusting the generator network parameters through the cross-granularity adversarial loss function;

[0046] The dynamic threshold feedback iteration module uses the dynamic behavior feature vectors of real users in the fine-grained category group as a benchmark, calculates the category matching score of the fine-grained marketing strategy candidate set through the local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix, and adopts the dynamic threshold adaptive strategy selection method to perform feedback iteration to update the generator network parameters, and finally determines the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

[0047] Optionally, the S4 specifically includes:

[0048] S41. Parameterize the fine-grained marketing strategy, define the strategy delivery duration parameter, discount range parameter, and push content combination parameter as decision variables, and construct a strategy parameter vector.

[0049] S42. Establish an outer layer Flying Fox optimization algorithm with the strategy parameter vector as the prey position and the overall user response trend change as the prey trajectory. The outer layer Flying Fox optimization algorithm uses the overall user response trend fitness function to calculate the overall fitness of the strategy parameter vector in real time:

[0050]

[0051] in, represents the strategy parameter vector of the outer flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, represents the current global optimal strategy parameter vector calculated based on the overall real-time user response trend change. Represents the dynamic inertia speed of the t-th outer strategy parameter update:

[0052]

[0053] Among them, ω is the dynamic inertia weight factor, α, β, and γ represent the inertia update weight, trend guidance weight, and user group response acceleration weight, respectively;

[0054] S43, receiving the strategy parameter vector output by the outer Feihu optimization algorithm with the inner Feihu optimization algorithm, and defining a local strategy response fitness function according to the optimization direction determined by the strategy parameter vector, for calculating the fitness value after local adjustment of each strategy parameter;

[0055] S44. Locally adjust the inner layer Flying Fox optimization algorithm strategy parameters:

[0056]

[0057] in, represents the strategy parameter vector of the inner flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, Represents the current local fitness optimal position vector, Represents the local space search enhancement factor:

[0058]

[0059] Where f(·) is the current local fitness function, ε is a very small positive number to prevent division by zero, η and λ represent the local adjustment speed weight and random oscillation adjustment weight, respectively, rand(0,1) represents a random number between 0 and 1, and σ is the oscillation intensity adjustment factor;

[0060] S45. Based on the optimal strategy parameter vector obtained after the inner flying fox group completes the iteration, the final marketing strategy parameter combination is determined and the fine-grained marketing strategy is output.

[0061] Optionally, the S6 specifically includes:

[0062] S61. Real-time monitoring of actual user response data after deployment of the fine-grained marketing strategy, and determining a strategy response difference vector based on difference data between the actual user response data and the expected response data;

[0063] S62: Identify feature dimensions with significant strategy response differences based on the strategy response difference vector, and determine parameter dimensions of the R3GAN strategy generation model that need to be adjusted first in the next round of marketing strategy generation;

[0064] S63, determining the real-time adjustment ratio of each parameter dimension in the R3GAN strategy generation model based on the difference amplitude of the strategy response difference vector;

[0065] S64, actively adjusting the generator network parameter space in the R3GAN strategy generation model in real time based on the determined real-time adjustment ratio, so that the generator network parameter space is synchronized with the real-time change trend of the actual user response data;

[0066] S65. Regenerate the fine-grained marketing strategy using the adjusted generator network parameters, and perform a strategy matching score on the regenerated fine-grained marketing strategy to obtain an updated strategy matching score;

[0067] S66. When the updated strategy matching score does not reach the preset adaptation threshold, repeat steps S62 to S65 until the strategy matching score meets the preset adaptation threshold, completing the real-time self-perception adaptation adjustment of the parameter space of the R3GAN strategy generation model.

[0068] Optionally, the S7 specifically includes:

[0069] S71. Define a marketing strategy performance decay threshold based on the strategy matching score sequence of historical fine-grained marketing strategies, and use the performance decay threshold as a criterion for determining whether to continue deploying the current marketing strategy.

[0070] S72. Real-time monitoring of actual user response data of the currently deployed fine-grained marketing strategy within a predetermined period, and calculation of a strategy matching score;

[0071] S73. Continuously calculate the strategy matching scores of the fine-grained marketing strategies, and compare the latest strategy matching scores with the set performance decay threshold in real time to determine whether the performance of the current marketing strategy is continuously below the performance decay threshold;

[0072] S74. When the strategy matching score of a marketing strategy falls below the performance attenuation threshold for multiple consecutive times, the marketing strategy retirement mechanism is automatically activated to decommission the currently deployed fine-grained marketing strategy in real time.

[0073] S75. When the current marketing strategy is retired, steps S3 to S6 are re-triggered to generate a new alternative fine-grained marketing strategy using the R3GAN strategy generation model that has undergone real-time self-perception adaptation and adjustment.

[0074] S76. The strategy parameters of the newly generated alternative fine-grained marketing strategy are optimized and adjusted through the Feihu optimization algorithm with a double-layer nested structure, and finally deployed on the e-commerce platform to replace the retired marketing strategy.

[0075] The beneficial effects of the present invention are:

[0076] (1) By constructing a phased and progressive R3GAN strategy generation model, the present invention realizes the gradual generation and optimization of coarse-grained marketing strategies and fine-grained marketing strategies, effectively improving the personalization and real-time adaptation capabilities of marketing strategies, and enhancing the e-commerce platform's ability to quickly respond to changes in user behavior.

[0077] (2) The present invention adopts the Feihu optimization algorithm with a double-layer nested structure, which can realize real-time dynamic optimization of marketing strategy parameters, significantly improving the optimization speed and accuracy of marketing strategies in terms of delivery duration, discount range and push content combination, and showing better adaptability in scenarios where users' real-time behavior changes.

[0078] (3) In terms of dynamic self-adaptive adjustment of the marketing strategy parameter space, the present invention actively adjusts the parameter space of the R3GAN strategy generation model in real time through the strategy self-perception adaptation mechanism, effectively solving the problem of slow response to real-time changes in user needs in the existing technology, breaking through the limitation of the traditional strategy model's static parameter space being difficult to adjust in real time, and realizing dynamic and accurate adaptation of marketing strategies to user needs, thereby effectively improving the marketing strategy recommendation effect of the e-commerce platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0080] Figure 1 This is a flow chart of the method for generating marketing strategies for e-commerce platforms based on generative adversarial networks proposed in the present invention. DETAILED DESCRIPTION

[0081] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0082] refer to Figure 1 , the e-commerce platform marketing strategy generation method based on generative adversarial network includes the following steps:

[0083] S1. Collect user behavior data on the e-commerce platform, pre-process the behavior data, and construct a user dynamic behavior data sequence based on time series;

[0084] S2. Use the R3GAN model to perform semi-supervised hierarchical clustering on the user dynamic behavior data sequence to form a hierarchical cluster group with features that change dynamically over time;

[0085] S3. Based on the hierarchical clustering of features, a phased and progressive R3GAN strategy generation model is constructed to generate coarse-grained marketing strategies and fine-grained marketing strategies respectively;

[0086] S4. Parameterize the fine-grained marketing strategy, construct a double-layer nested Feihu optimization algorithm, and optimize the fine-grained marketing strategy;

[0087] S5. Deploy the optimized fine-grained marketing strategy in the e-commerce platform operating environment, monitor actual user response data in real time, and record the difference between the user's real-time response and the expected response;

[0088] S6. Through the strategy self-perception adaptation mechanism, the parameter space of the phased progressive R3GAN strategy generation model is actively adjusted in real time based on the difference data, so that the next round of marketing strategy generation process can autonomously adapt to the real-time user response changes;

[0089] S7. Set a performance attenuation threshold for the marketing strategy and monitor the performance of the marketing strategy in real time. When the sustained performance falls below the attenuation threshold, the retirement mechanism of the marketing strategy is automatically triggered, and the process of steps S3 to S6 is immediately restarted to generate and deploy a new replacement marketing strategy.

[0090] By adopting a phased and progressive R3GAN strategy generation model and combining it with the Feihu optimization algorithm with a double-layer nested structure, we have achieved refined parameter optimization and real-time dynamic adaptation of marketing strategies, which can promptly capture and respond to changes in user needs; at the same time, we have constructed a strategy self-perception adaptation mechanism and a performance degradation automatic retirement mechanism, which effectively avoids the problem of continuous poor performance of marketing strategies and improves the accuracy of marketing strategies and user response efficiency.

[0091] In this embodiment, S1 specifically includes:

[0092] S11. During each page visit in the user's browsing history, the page scroll position changes and mouse movement trajectory are captured in real time, and combined with the page dwell time, page focus point trajectory data is established;

[0093] S12. Identify and extract the user's search intent change process based on the user's search keywords and the secondary modified search behavior after each search, and construct search intent evolution data reflecting the continuous evolution of search intent;

[0094] S13. For each click event in the number of user clicks, extract the positional relationship between the click and the page element and the page element attributes to obtain click spatial feature data of the click action;

[0095] S14. Extracting dynamic hesitation behavior data of the user's shopping decision based on the time interval between adding, removing, modifying the quantity of the shopping cart, and final purchase completion in the user's purchase behavior to form purchase hesitation behavior data;

[0096] S15. Based on the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data, identify and filter out abnormal patterns in temporal features, retaining the normal fluctuation characteristics of dynamic behavior as the user behavior dynamic sequence;

[0097] S16. Using the historical variation ranges of the user behavior dynamic sequences, respectively set the normalized upper and lower limits of dynamic adjustment to construct a user dynamic behavior data sequence that reflects the dynamic trends of the user's real behavior.

[0098] By capturing the user's page focus trajectory, search intent evolution, click space characteristics and purchase decision dynamic hesitation behavior data in real time, and combining it with the recognition of abnormal patterns of time series characteristics and dynamic adjustment of the normalization range, the dynamics, accuracy and detail of the user behavior data sequence are effectively improved, thereby improving the generation quality of subsequent marketing strategies and user adaptation effects.

[0099] In this embodiment, S2 specifically includes:

[0100] S21. Align the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data in the user dynamic behavior data sequence according to timestamps to form a time-synchronized dynamic behavior feature vector.

[0101] S22. Randomly select a preset proportion of dynamic behavior feature vectors that have been labeled with user interest categories as the supervision sample set in the clustering process, and the dynamic behavior feature vectors that have not been labeled with user interest categories as the sample set to be classified;

[0102] S23. Using the generator in the R3GAN model, with the dynamic behavior feature vectors and corresponding interest category labels of the supervised sample set as the benchmark, multiple parallel generation channels are constructed. Each channel introduces Gaussian random noise of different variances through the feature noise input layer, and generates interest category prediction labels for the sample set to be classified channel by channel. The label difference distribution between channels is used to dynamically adjust the output weights of each channel in the generator, forming a preliminary category prediction result with a difference enhancement effect.

[0103] S24. Using the discriminator in the R3GAN model, the dynamic behavior feature vectors of the supervised sample set that have been labeled with the interest category label are used as positive samples, and the dynamic behavior feature vectors of the sample set to be classified corresponding to the preliminary category prediction results are used as negative samples. The local feature differences and global feature differences of the corresponding categories of the positive and negative samples are calculated respectively, and the sum of the weighted local feature differences and the global feature differences is used as the category matching score. Based on the category matching score, the category boundary is dynamically corrected through the adaptive threshold iteration method, and the final category matching score is output.

[0104] S25. Determine the final category label of the sample set to be classified based on the final category matching score, and complete the preliminary clustering of the dynamic behavior feature vector in combination with the supervision sample set;

[0105] S26. Based on the continuous change characteristics of the dynamic behavior feature vector on the time axis, the categories of the preliminary clustering are dynamically stratified to form feature-stratified cluster groups that change dynamically over time.

[0106] By utilizing the generator and discriminator in the R3GAN model, a multi-channel interest category prediction mechanism and a dynamically corrected category matching scoring method are constructed. This enables the clustering process of users' dynamic behavior characteristics to accurately identify and adapt to the dynamic evolution of user interests, effectively improving the personalization level and dynamic adaptability of marketing strategy generation, thereby optimizing marketing effectiveness.

[0107] In this embodiment, S3 specifically includes:

[0108] S31, setting a granularity classification standard for the feature hierarchical cluster group according to the number of samples and the degree of dynamic behavior difference of different feature categories in the feature hierarchical cluster group, and dividing the cluster group into a coarse-grained category group and a fine-grained category group;

[0109] S32. Using the dynamic behavior feature vectors within the coarse-grained category groups as training samples, the generator network of the first-stage R3GAN strategy generation model is constructed. By introducing multiple groups of high-dimensional random noise vectors with different variances into the input layer of the dynamic behavior feature vectors, the initial coarse-grained marketing strategy candidate set is generated based on the dynamic combination of the transformation factors of each random noise.

[0110] S33, using the discriminator network of the first-stage R3GAN strategy generation model to determine a coarse-grained marketing strategy that adapts to the initial user demand trend;

[0111] S34. Subdivide the coarse-grained marketing strategy into a set of strategy feature sub-vectors through nonlinear mapping of the strategy feature dimension, which serves as the input benchmark of the second-stage R3GAN strategy generation model, and uses the dynamic behavior feature vectors within the fine-grained category group as supervision samples;

[0112] S35. In the generator network of the second-stage R3GAN strategy generation model, local perturbation random noise with gradually decreasing scale is introduced on each feature dimension of the strategy feature sub-vector set, gradually generating a number of fine-grained marketing strategy candidate sets with differentiated feature details;

[0113] S36. Through the discriminator network of the second-stage R3GAN strategy generation model, the real user dynamic behavior feature vectors in the fine-grained category group are used as positive samples, and the generated fine-grained marketing strategy candidate set is used as negative samples. The local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix between the positive and negative samples are calculated respectively, and the Frobenius norm of the two difference distribution matrices is used as the scoring indicator of the strategy category matching degree. Then, feedback iteration is performed based on the score gradient descent method to finally determine the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

[0114] By adopting a phased and progressive R3GAN strategy generation model, we first use a coarse-grained strategy to preliminarily match user demand trends, and then combine it with a local perturbation mechanism with gradually decreasing scale to generate a fine-grained strategy. This effectively solves the problem that the existing single-stage strategy model is not sufficiently responsive to changes in user demand, and significantly improves the adaptation accuracy and flexibility of marketing strategies and changes in user personalized trends.

[0115] In this embodiment, the first-stage R3GAN strategy generation model specifically includes a feature adaptive enhancement input module, a random noise multi-scale interaction generation module, a strategy time-space collaborative fusion module, a dynamic strategy self-confrontation module, and an adaptive momentum feedback iteration module:

[0116] The feature adaptive enhancement input module is used to receive the dynamic behavior feature vector within the coarse-grained category group, and based on the internal distance difference of the feature vector space, adaptively amplify the feature value of the key feature dimension within the dynamic behavior feature vector through the feature enhancement factor;

[0117] The random noise multi-scale interactive generation module uses multi-scale Gaussian random noise and generates multiple groups of scale interactive random noise vectors through the scale interactive fusion process:

[0118]

[0119] Among them, Z i represents the interactive fusion random noise vector of the i-th scale, represents the variance of random noise at the i-th scale, represents the variance of the random noise at the jth scale, λ ij is the noise interaction fusion coefficient between scales, K represents the total number of random noise scales;

[0120] The formula introduces Gaussian random noise of different scales in the random noise generation process and interactively fuses multiple scales to achieve a richer and more diverse random feature representation. Among them, the interactive fusion random noise vector Z of each scale is i It is formed by weighted superposition of random noise of this scale and random noise of other scales. Random noise of different scales is fused by interaction coefficient λ ij They interact with each other, thereby effectively enhancing the feature diversity and stability of the generated data and improving the strategy generation performance of the generative adversarial model in complex feature spaces.

[0121] The strategic temporal-spatial collaborative fusion module inputs the enhanced features of the feature adaptive enhancement input module and the scale-interactive random noise vector generated by the random noise multi-scale interaction generation module, and forms a collaborative fusion process of strategic features by respectively constructing a temporal attention mechanism and a spatial attention mechanism, including generating temporal attention weights through a bidirectional gated recurrent unit network and generating spatial attention weights through a two-dimensional convolutional network, and then gradually fusing the temporal attention weights and the spatial attention weights to obtain a fused feature expression;

[0122] The dynamic strategy self-adversarial module generates an initial coarse-grained marketing strategy candidate set based on the fusion feature expression, randomly groups the strategies in the coarse-grained marketing strategy candidate set, and conducts self-adversarial training on the strategy features between each two. The generator network parameters are dynamically adjusted through the adversarial loss function between the alternative strategies:

[0123] L adv =∑ m,n∈G,m≠n max(0,∈-d(X m ,X n ));

[0124] Among them, G is the strategy alternative set, X m 、X n are the feature vectors of alternative strategies, d(·) is the feature distance function between alternative strategies, and ∈ is the minimum distance threshold between strategies;

[0125] The formula implements dynamic strategy self-adversarial training by randomly grouping each strategy in the coarse-grained marketing strategy alternative set and calculating the feature distance between strategies. adv By using the difference between the minimum distance threshold ∈ between strategies and the actual feature distance, the feature vectors of each strategy are encouraged to maintain sufficiently large differences from each other.

[0126] The adaptive momentum feedback iteration module calculates the matching scores between the strategies in the marketing strategy alternative set and the real features based on the real user dynamic behavior feature vector, adaptively adjusts the momentum parameters according to the score gradient change, dynamically updates the generator network parameters, and outputs a coarse-grained marketing strategy that ultimately adapts to the initial user demand trend.

[0127] Through the feature adaptive enhanced input, multi-scale interactive fusion of random noise and time-space collaborative fusion technology of strategy features, efficient and diversified generation of coarse-grained marketing strategies is achieved. And through the dynamic strategy self-adversarial training mechanism and adaptive momentum feedback iteration method, the ability to accurately match the generated strategy with user demand trends is further enhanced, effectively improving the stability and adaptation accuracy of marketing strategy generation.

[0128] In this embodiment, the second-stage R3GAN strategy generation model specifically includes a strategy feature fine-grained recoding module, a strategy dimension step-by-step perturbation module, a multi-head adaptive attention fusion module, a cross-granularity strategy feature adversarial module, and a dynamic threshold feedback iteration module:

[0129] The strategy feature fine-grained recoding module is used to receive a set of strategy feature sub-vectors and, based on a feature difference matrix of dynamic behavior feature vectors within a fine-grained category group, perform dimension-by-dimension adaptive fine-grained recoding on the strategy feature sub-vectors through feature recoding to obtain fine-grained strategy initial coding features;

[0130] The strategy dimension step-by-step perturbation module adopts Gaussian random noise with decreasing scale, and embeds the fine-grained strategy initial encoding features dimension by dimension through the dimension perturbation factor to obtain the dimension perturbation feature:

[0131]

[0132] Among them, X′ v Represents the feature dimension vector after perturbation, X v Represents the v-th dimension feature value in the initial encoding feature of the fine-grained strategy, μ v is the perturbation factor of the vth dimension, represents the variance of random noise in the vth dimension, and M is the total number of feature dimensions;

[0133] The formula achieves refined perturbation optimization of the strategy features in each dimension by embedding Gaussian random noise with decreasing scale in each dimension in the initial encoding features of the fine-grained strategy. Among them, the dimension perturbation feature vector X′ v From the original eigenvalue X v With the perturbation factor μ v The strategy feature space is explored in a step-by-step refined manner, thereby improving the sensitivity and adaptation accuracy of the generated fine-grained strategy to changes in user personalization trends.

[0134] The multi-head adaptive attention fusion module receives dimensional perturbation features, extracts strategy temporal attention weights, dimensional feature importance weights, and feature interaction weights by constructing a multi-head adaptive attention network structure, and dynamically fuses them according to the weight vector set to obtain a multi-head fusion feature expression;

[0135] The cross-granularity strategy feature adversarial module generates a fine-grained marketing strategy candidate set by expressing multi-head fusion features, and performs cross-granularity feature adversarial training with the coarse-grained marketing strategy output by the first-stage R3GAN strategy generation model. The generator network parameters are dynamically adjusted through the cross-granularity adversarial loss function:

[0136] L cross-adv =∑p∈F,q∈C max(0,δ-d(X p ,X q ));

[0137] Among them, F represents the candidate set of fine-grained marketing strategies, C represents the coarse-grained marketing strategies, and X p 、X q denote the fine-grained and coarse-grained strategy feature vectors, d(·) is the strategy feature distance function, and δ is the minimum distance threshold across the strategies;

[0138] The formula calculates the feature distance between fine-grained and coarse-grained marketing strategy candidate sets through the cross-granularity strategy feature adversarial training method, and uses the cross-granularity strategy minimum distance threshold δ as the benchmark to encourage significant feature differences between strategies. cross-adv By widening the distance between the characteristics of strategies of different granularities, fine-grained marketing strategies can effectively inherit the overall trend of coarse-grained strategies.

[0139] The dynamic threshold feedback iteration module uses the dynamic behavior feature vectors of real users in the fine-grained category group as a benchmark, calculates the category matching score of the fine-grained marketing strategy candidate set through the local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix, and adopts the dynamic threshold adaptive strategy selection method to perform feedback iteration to update the generator network parameters, and finally determines the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

[0140] Through fine-grained feature recoding, multi-head adaptive attention fusion and cross-granularity strategy feature adversarial training mechanism, the refinement and dynamic optimization of fine-grained marketing strategy features are achieved, and the model parameters are iteratively adjusted through dynamic threshold feedback, which effectively improves the generation strategy's ability to accurately adapt to subtle changes in user personalized trends, further enhancing the accuracy and stability of the marketing strategy.

[0141] In this embodiment, the S4 specifically includes:

[0142] S41. Parameterize the fine-grained marketing strategy, define the strategy delivery duration parameter, discount range parameter, and push content combination parameter as decision variables, and construct a strategy parameter vector.

[0143] S42. Establish an outer layer Flying Fox optimization algorithm with the strategy parameter vector as the prey position and the overall user response trend change as the prey trajectory. The outer layer Flying Fox optimization algorithm uses the overall user response trend fitness function to calculate the overall fitness of the strategy parameter vector in real time:

[0144]

[0145] in, represents the strategy parameter vector of the outer flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, represents the current global optimal strategy parameter vector calculated based on the overall real-time user response trend change. Represents the dynamic inertia speed of the t-th outer strategy parameter update:

[0146]

[0147] Among them, ω is the dynamic inertia weight factor, α, β, and γ represent the inertia update weight, trend guidance weight, and user group response acceleration weight, respectively;

[0148] The formula uses the outer layer Feihu optimization algorithm to achieve dynamic updates of the overall strategy parameters, where the strategy parameter vector update process integrates the inertia speed, the optimal direction of the overall user response trend, and the acceleration effect of the user group response change. From the previous round parameter vector and the current trend optimal parameter vector Combined, while introducing dynamic inertia speed Through a smooth update process, marketing strategy parameters can be effectively ensured to quickly and stably follow changes in overall user demand trends, thereby enhancing the response speed and accuracy of overall parameter optimization.

[0149] S43, receiving the strategy parameter vector output by the outer Feihu optimization algorithm with the inner Feihu optimization algorithm, and defining a local strategy response fitness function according to the optimization direction determined by the strategy parameter vector, for calculating the fitness value after local adjustment of each strategy parameter;

[0150] S44. Locally adjust the inner layer Flying Fox optimization algorithm strategy parameters:

[0151]

[0152] in, represents the strategy parameter vector of the inner flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, Represents the current local fitness optimal position vector, Represents the local space search enhancement factor:

[0153]

[0154] Where f(·) is the current local fitness function, ε is a very small positive number to prevent division by zero, η and λ represent the local adjustment speed weight and random oscillation adjustment weight, respectively, rand(0,1) represents a random number between 0 and 1, and σ is the oscillation intensity adjustment factor;

[0155] The formula achieves efficient fine-tuning of the strategy parameter vector through local fine-tuning of the inner Feihu optimization algorithm. The update process of the strategy parameter vector takes into account the difference between the current optimal local position and the current parameter position, the random oscillation term, and the local space search enhancement factor. By dynamically adjusting the search step size based on the relative differences in the local fitness of the current parameters, the sensitivity of parameter fine-tuning and optimization efficiency are effectively improved, and the ability of the strategy parameters to converge quickly and accurately within a local range is enhanced.

[0156] S45. Based on the optimal strategy parameter vector obtained after the inner flying fox group completes the iteration, the final marketing strategy parameter combination is determined and the fine-grained marketing strategy is output.

[0157] By constructing a double-layer nested Feihu optimization algorithm, the outer Feihu optimization algorithm tracks the overall user response trend in real time to determine the global parameter optimization direction, and the inner Feihu optimization algorithm quickly and finely adjusts the strategy parameters according to the optimization direction determined by the outer layer, effectively improving the response speed and accuracy of marketing strategy parameter optimization, and significantly improving the strategy's adaptability and actual marketing performance in a dynamic user demand environment.

[0158] In this embodiment, S6 specifically includes:

[0159] S61. Real-time monitoring of actual user response data after deployment of the fine-grained marketing strategy, and determining a strategy response difference vector based on difference data between the actual user response data and the expected response data;

[0160] S62: Identify feature dimensions with significant strategy response differences based on the strategy response difference vector, and determine parameter dimensions of the R3GAN strategy generation model that need to be adjusted first in the next round of marketing strategy generation;

[0161] S63, determining the real-time adjustment ratio of each parameter dimension in the R3GAN strategy generation model based on the difference amplitude of the strategy response difference vector;

[0162] S64, actively adjusting the generator network parameter space in the R3GAN strategy generation model in real time based on the determined real-time adjustment ratio, so that the generator network parameter space is synchronized with the real-time change trend of the actual user response data;

[0163] S65. Regenerate the fine-grained marketing strategy using the adjusted generator network parameters, and perform a strategy matching score on the regenerated fine-grained marketing strategy to obtain an updated strategy matching score;

[0164] S66. When the updated strategy matching score does not reach the preset adaptation threshold, repeat steps S62 to S65 until the strategy matching score meets the preset adaptation threshold, completing the real-time self-perception adaptation adjustment of the parameter space of the R3GAN strategy generation model.

[0165] By building a strategy self-perception adaptation mechanism, monitoring the actual user response after the marketing strategy is deployed in real time, and actively adjusting the parameter space of the R3GAN strategy generation model in real time, efficient real-time adaptation of the generation model parameters and user response trends is achieved, ensuring the continuous optimization of the marketing strategy, and effectively improving the real-time adaptability of strategy generation and the matching accuracy of user responses.

[0166] In this embodiment, the S7 specifically includes:

[0167] S71. Define a marketing strategy performance decay threshold based on the strategy matching score sequence of historical fine-grained marketing strategies, and use the performance decay threshold as a criterion for determining whether to continue deploying the current marketing strategy.

[0168] S72. Real-time monitoring of actual user response data of the currently deployed fine-grained marketing strategy within a predetermined period, and calculation of a strategy matching score;

[0169] S73. Continuously calculate the strategy matching scores of the fine-grained marketing strategies, and compare the latest strategy matching scores with the set performance decay threshold in real time to determine whether the performance of the current marketing strategy is continuously below the performance decay threshold;

[0170] S74. When the strategy matching score of a marketing strategy falls below the performance attenuation threshold for multiple consecutive times, the marketing strategy retirement mechanism is automatically activated to decommission the currently deployed fine-grained marketing strategy in real time.

[0171] S75. When the current marketing strategy is retired, steps S3 to S6 are re-triggered to generate a new alternative fine-grained marketing strategy using the R3GAN strategy generation model that has undergone real-time self-perception adaptation and adjustment.

[0172] S76. The strategy parameters of the newly generated alternative fine-grained marketing strategy are optimized and adjusted through the Feihu optimization algorithm with a double-layer nested structure, and finally deployed on the e-commerce platform to replace the retired marketing strategy.

[0173] By setting a performance attenuation threshold for marketing strategies and implementing real-time monitoring and automatic retirement mechanisms, when a strategy's performance continues to fall below the performance threshold, it will be immediately taken offline and the strategy generation process restarted to ensure that the marketing strategy is always in a high-performance state, effectively avoiding the problem of long-term inefficient operation of the strategy and improving the overall adaptability and continuous optimization capabilities of the marketing strategy.

[0174] Example 1:

[0175] In order to verify the feasibility of the present invention in implementation, the present invention is applied to the real-time generation task of intelligent marketing strategies of a certain e-commerce platform to improve the response speed and accuracy of marketing strategies to changes in user needs. The platform previously adopted the traditional static marketing strategy generation method, which mainly established a unified model based on historical user data, and then updated the marketing content through a fixed period. However, due to the low frequency of data updates, the traditional solution is unable to capture the real-time changing trends of user behavior in a timely manner, resulting in insufficient personalization of marketing content and low user interaction rate, which in turn affects the overall marketing conversion effect. In addition, the traditional method uses a single algorithm, which makes it difficult to achieve efficient coordination between marketing content and parameter optimization, resulting in slow response to strategy adjustments and large differences in user responses.

[0176] In practical applications, the system first collects multi-dimensional behavioral data on users' browsing history, search history, click counts, purchase history, and page dwell time on e-commerce platforms in real time. Through data denoising and normalization, it constructs a dynamic behavioral data sequence that reflects users' real-time interest trends. This dynamic behavioral data is then fed into the semi-supervised hierarchical clustering R3GAN model. Based on the similarity of data features and temporal continuity, the system divides the user group into hierarchical clusters of different interest categories in real time, thereby more accurately describing the dynamic changes in users' interest preferences.

[0177] Next, based on the division of user behavior clusters, a phased and progressive R3GAN strategy generation model was constructed. This model first used preliminary clustering results to generate coarse-grained marketing strategies to capture overall trends. Then, through strategy dimension segmentation and feature perturbation methods, a more refined and personalized fine-grained marketing strategy was generated. To ensure the real-time accuracy of the marketing strategy, a two-layer nested Feihu optimization algorithm was further used for dynamic parameter optimization. The outer Feihu algorithm tracked the response trends of the entire user base in real time to determine the optimization direction, while the inner Feihu algorithm rapidly adjusted the specific parameters of the marketing strategy, such as the duration of the campaign, the strength of the discount, and the combination of push content.

[0178] After the marketing strategy is actually deployed, the strategy effectiveness is evaluated through real-time user response data, including key indicators such as click-through rate, interaction rate and conversion rate. Based on the difference between the actual response and the expected response, a strategy self-perception adaptation mechanism is constructed to achieve active and real-time adjustment of the R3GAN model parameter space, ensuring that the next round of strategy generation can autonomously and dynamically adapt to the real-time changes in user needs.

[0179] In addition, we've set performance decay thresholds for marketing strategies to continuously monitor their real-time performance. When a strategy's performance continuously falls below the set decay threshold, the system automatically triggers its retirement, quickly replacing it with a new replacement strategy to maintain the high quality and efficiency of marketing content.

[0180] During the actual operation, the marketing data for one month was monitored and analyzed to verify the effectiveness of the method of the present invention. The data table is shown as follows:

[0181] Table 1: Comparison of marketing strategy performance before and after implementing the method of the present invention

[0182]

[0183] As can be seen from the data in Table 1, after the implementation of the present invention, key indicators such as the click-through rate, conversion rate, and interaction rate of the marketing strategy have all been significantly improved. Among them, the click-through rate increased from 2.5% before implementation to 4.3%, an increase of 72.0%; the conversion rate increased from the original 1.1% to 2.6%, an increase of 136.4%; and the interaction rate increased by more than double, from 4.2% to 8.9%. At the same time, the dynamic optimization and adaptive mechanism of the present invention significantly shortened the average response adjustment time from 48 hours to 0.5 hours, an increase of more than 95%, greatly improving the efficiency of the marketing strategy in responding to real-time changes in user needs.

[0184] The data results of this embodiment fully verify the technical advantages of the present invention in real-time personalized marketing strategy generation and dynamic optimization and adjustment, effectively solve the problems of insufficient user demand adaptation and delayed strategy response in traditional methods, significantly improve the marketing efficiency and user experience of e-commerce platforms, and provide more accurate, efficient and dynamic technical solutions for real-time marketing scenarios, reflecting significant economic benefits and technological advancement, and has broad application prospects and promotion value.

[0185] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The e-commerce platform marketing strategy generation method based on generative adversarial network is characterized by: The steps include: S1. Collect user behavior data on the e-commerce platform, pre-process the behavior data, and construct a user dynamic behavior data sequence based on time series; S2. Use the R3GAN model to perform semi-supervised hierarchical clustering on the user dynamic behavior data sequence to form a hierarchical cluster group with features that change dynamically over time; S3. Based on the hierarchical clustering of features, a phased and progressive R3GAN strategy generation model is constructed to generate coarse-grained marketing strategies and fine-grained marketing strategies respectively; S4. Parameterize the fine-grained marketing strategy, construct a double-layer nested Feihu optimization algorithm, and optimize the fine-grained marketing strategy; S5. Deploy the optimized fine-grained marketing strategy in the e-commerce platform operating environment, monitor actual user response data in real time, and record the difference between the user's real-time response and the expected response; S6. Through the strategy self-perception adaptation mechanism, the parameter space of the phased progressive R3GAN strategy generation model is actively adjusted in real time based on the difference data, so that the next round of marketing strategy generation process can autonomously adapt to the real-time user response changes; S7. Set a performance attenuation threshold for the marketing strategy and monitor the performance of the marketing strategy in real time. When the sustained performance falls below the attenuation threshold, the retirement mechanism of the marketing strategy is automatically triggered, and the process of steps S3 to S6 is immediately restarted to generate and deploy a new replacement marketing strategy.

2. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: Said S1 specifically includes: S11. During each page visit in the user's browsing history, the page scroll position changes and mouse movement trajectory are captured in real time, and combined with the page dwell time, page focus point trajectory data is established; S12. Identify and extract the user's search intent change process based on the user's search keywords and the secondary modified search behavior after each search, and construct search intent evolution data reflecting the continuous evolution of search intent; S13. For each click event in the number of user clicks, extract the positional relationship between the click and the page element and the page element attributes to obtain click spatial feature data of the click action; S14. Extracting dynamic hesitation behavior data of the user's shopping decision based on the time interval between adding, removing, modifying the quantity of the shopping cart, and final purchase completion in the user's purchase behavior to form purchase hesitation behavior data; S15. Based on the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data, identify and filter out abnormal patterns in temporal features, retaining the normal fluctuation characteristics of dynamic behavior as the user behavior dynamic sequence; S16. Using the historical variation ranges of the user behavior dynamic sequences, respectively set the normalized upper and lower limits of dynamic adjustment to construct a user dynamic behavior data sequence that reflects the dynamic trends of the user's real behavior.

3. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: The S2 specifically includes: S21. Align the page focus trajectory data, search intent evolution data, click space feature data, and purchase hesitation behavior data in the user dynamic behavior data sequence according to timestamps to form a time-synchronized dynamic behavior feature vector. S22. Randomly select a preset proportion of dynamic behavior feature vectors that have been labeled with user interest categories as the supervision sample set in the clustering process, and the dynamic behavior feature vectors that have not been labeled with user interest categories as the sample set to be classified; S23. Using the generator in the R3GAN model, with the dynamic behavior feature vectors and corresponding interest category labels of the supervised sample set as the benchmark, multiple parallel generation channels are constructed. Each channel introduces Gaussian random noise of different variances through the feature noise input layer, and generates interest category prediction labels for the sample set to be classified channel by channel. The label difference distribution between channels is used to dynamically adjust the output weights of each channel in the generator, forming a preliminary category prediction result with a difference enhancement effect. S24. Using the discriminator in the R3GAN model, the dynamic behavior feature vectors of the supervised sample set that have been labeled with the interest category label are used as positive samples, and the dynamic behavior feature vectors of the sample set to be classified corresponding to the preliminary category prediction results are used as negative samples. The local feature differences and global feature differences of the corresponding categories of the positive and negative samples are calculated respectively, and the sum of the weighted local feature differences and the global feature differences is used as the category matching score. Based on the category matching score, the category boundary is dynamically corrected through the adaptive threshold iteration method, and the final category matching score is output. S25. Determine the final category label of the sample set to be classified based on the final category matching score, and complete the preliminary clustering of the dynamic behavior feature vector in combination with the supervision sample set; S26. Based on the continuous change characteristics of the dynamic behavior feature vector on the time axis, the categories of the preliminary clustering are dynamically stratified to form feature-stratified cluster groups that change dynamically over time.

4. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: The S3 specifically includes: S31, setting a granularity classification standard for the feature hierarchical cluster group according to the number of samples and the degree of dynamic behavior difference of different feature categories in the feature hierarchical cluster group, and dividing the cluster group into a coarse-grained category group and a fine-grained category group; S32. Using the dynamic behavior feature vectors within the coarse-grained category groups as training samples, the generator network of the first-stage R3GAN strategy generation model is constructed. By introducing multiple groups of high-dimensional random noise vectors with different variances into the input layer of the dynamic behavior feature vectors, the initial coarse-grained marketing strategy candidate set is generated based on the dynamic combination of the transformation factors of each random noise. S33, using the discriminator network of the first-stage R3GAN strategy generation model to determine a coarse-grained marketing strategy that adapts to the initial user demand trend; S34. Subdivide the coarse-grained marketing strategy into a set of strategy feature sub-vectors through nonlinear mapping of the strategy feature dimension, which serves as the input benchmark of the second-stage R3GAN strategy generation model, and uses the dynamic behavior feature vectors within the fine-grained category group as supervision samples; S35. In the generator network of the second-stage R3GAN strategy generation model, local perturbation random noise with gradually decreasing scale is introduced on each feature dimension of the strategy feature sub-vector set, gradually generating a number of fine-grained marketing strategy candidate sets with differentiated feature details; S36. Through the discriminator network of the second-stage R3GAN strategy generation model, the real user dynamic behavior feature vectors in the fine-grained category group are used as positive samples, and the generated fine-grained marketing strategy candidate set is used as negative samples. The local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix between the positive and negative samples are calculated respectively, and the Frobenius norm of the two difference distribution matrices is used as the scoring indicator of the strategy category matching degree. Then, feedback iteration is performed based on the score gradient descent method to finally determine the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

5. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 4, characterized in that: The first-stage R3GAN strategy generation model specifically includes a feature adaptive enhancement input module, a random noise multi-scale interaction generation module, a strategy time-space collaborative fusion module, a dynamic strategy self-confrontation module, and an adaptive momentum feedback iteration module: The feature adaptive enhancement input module is used to receive the dynamic behavior feature vector within the coarse-grained category group, and based on the internal distance difference of the feature vector space, adaptively amplify the feature value of the key feature dimension within the dynamic behavior feature vector through the feature enhancement factor; The random noise multi-scale interactive generation module uses multi-scale Gaussian random noise and generates multiple groups of scale interactive random noise vectors through an inter-scale interactive fusion process; The strategic temporal-spatial collaborative fusion module inputs the enhanced features of the feature adaptive enhancement input module and the scale-interactive random noise vector generated by the random noise multi-scale interaction generation module, and forms a collaborative fusion process of strategic features by respectively constructing a temporal attention mechanism and a spatial attention mechanism, including generating temporal attention weights through a bidirectional gated recurrent unit network and generating spatial attention weights through a two-dimensional convolutional network, and then gradually fusing the temporal attention weights and the spatial attention weights to obtain a fused feature expression; The dynamic strategy self-adversarial module generates an initial coarse-grained marketing strategy candidate set based on the fusion feature expression, randomly groups the strategies in the coarse-grained marketing strategy candidate set, performs self-adversarial training on the strategy features between each two, and dynamically adjusts the generator network parameters through the adversarial loss function between the alternative strategies; The adaptive momentum feedback iteration module calculates the matching scores between the strategies in the marketing strategy alternative set and the real features based on the real user dynamic behavior feature vector, adaptively adjusts the momentum parameters according to the score gradient change, dynamically updates the generator network parameters, and outputs a coarse-grained marketing strategy that ultimately adapts to the initial user demand trend.

6. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 4, characterized in that: The second-stage R3GAN strategy generation model specifically includes a strategy feature fine-grained recoding module, a strategy dimension step-by-step perturbation module, a multi-head adaptive attention fusion module, a cross-granularity strategy feature adversarial module, and a dynamic threshold feedback iteration module: The strategy feature fine-grained recoding module is used to receive a set of strategy feature sub-vectors and, based on a feature difference matrix of dynamic behavior feature vectors within a fine-grained category group, perform dimension-by-dimension adaptive fine-grained recoding on the strategy feature sub-vectors through feature recoding to obtain fine-grained strategy initial coding features; The strategy dimension step-by-step perturbation module adopts Gaussian random noise with decreasing scale, and embeds the fine-grained strategy initial coding features dimension by dimension through the dimension perturbation factor to obtain the dimension perturbation feature; The multi-head adaptive attention fusion module receives dimensional perturbation features, extracts strategy temporal attention weights, dimensional feature importance weights, and feature interaction weights by constructing a multi-head adaptive attention network structure, and dynamically fuses them according to the weight vector set to obtain a multi-head fusion feature expression; The cross-granularity strategy feature adversarial module generates a fine-grained marketing strategy candidate set from multi-head fusion feature expressions, and performs cross-granularity feature adversarial training with the coarse-grained marketing strategies output by the first-stage R3GAN strategy generation model, dynamically adjusting the generator network parameters through the cross-granularity adversarial loss function; The dynamic threshold feedback iteration module uses the dynamic behavior feature vectors of real users in the fine-grained category group as a benchmark, calculates the category matching score of the fine-grained marketing strategy candidate set through the local behavior feature difference distribution matrix and the global behavior feature difference distribution matrix, and adopts the dynamic threshold adaptive strategy selection method to perform feedback iteration to update the generator network parameters, and finally determines the fine-grained marketing strategy that adapts to the subtle trend changes of user personalization.

7. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: The S4 specifically includes: S41. Parameterize the fine-grained marketing strategy, define the strategy delivery duration parameter, discount range parameter, and push content combination parameter as decision variables, and construct a strategy parameter vector. S42. Establish an outer layer Flying Fox optimization algorithm with the strategy parameter vector as the prey position and the overall user response trend change as the prey trajectory. The outer layer Flying Fox optimization algorithm uses the overall user response trend fitness function to calculate the overall fitness of the strategy parameter vector in real time: in, represents the strategy parameter vector of the outer flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, represents the current global optimal strategy parameter vector calculated based on the overall real-time user response trend change. Represents the dynamic inertia speed of the t-th outer strategy parameter update: Among them, ω is the dynamic inertia weight factor, α, β, and γ represent the inertia update weight, trend guidance weight, and user group response acceleration weight, respectively; S43, receiving the strategy parameter vector output by the outer Feihu optimization algorithm with the inner Feihu optimization algorithm, and defining a local strategy response fitness function according to the optimization direction determined by the strategy parameter vector, for calculating the fitness value after local adjustment of each strategy parameter; S44. Locally adjust the inner layer Flying Fox optimization algorithm strategy parameters: in, represents the strategy parameter vector of the inner flying fox group at the tth iteration, is the strategy parameter vector after the t+1th iteration, Represents the current local fitness optimal position vector, Represents the local space search enhancement factor: Where f(·) is the current local fitness function, ε is a very small positive number to prevent division by zero, η and λ represent the local adjustment speed weight and random oscillation adjustment weight, respectively, rand(0,1) represents a random number between 0 and 1, and σ is the oscillation intensity adjustment factor; S45. Based on the optimal strategy parameter vector obtained after the inner flying fox group completes the iteration, the final marketing strategy parameter combination is determined and the fine-grained marketing strategy is output.

8. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: The S6 specifically includes: S61. Real-time monitoring of actual user response data after deployment of the fine-grained marketing strategy, and determining a strategy response difference vector based on difference data between the actual user response data and the expected response data; S62: Identify feature dimensions with significant strategy response differences based on the strategy response difference vector, and determine parameter dimensions of the R3GAN strategy generation model that need to be adjusted first in the next round of marketing strategy generation; S63, determining the real-time adjustment ratio of each parameter dimension in the R3GAN strategy generation model based on the difference amplitude of the strategy response difference vector; S64, actively adjusting the generator network parameter space in the R3GAN strategy generation model in real time based on the determined real-time adjustment ratio, so that the generator network parameter space is synchronized with the real-time change trend of the actual user response data; S65. Regenerate the fine-grained marketing strategy using the adjusted generator network parameters, and perform a strategy matching score on the regenerated fine-grained marketing strategy to obtain an updated strategy matching score; S66. When the updated strategy matching score does not reach the preset adaptation threshold, repeat steps S62 to S65 until the strategy matching score meets the preset adaptation threshold, completing the real-time self-perception adaptation adjustment of the parameter space of the R3GAN strategy generation model.

9. The method for generating marketing strategies for e-commerce platforms based on generative adversarial networks according to claim 1, characterized in that: The S7 specifically includes: S71. Define a marketing strategy performance decay threshold based on the strategy matching score sequence of historical fine-grained marketing strategies, and use the performance decay threshold as a criterion for determining whether to continue deploying the current marketing strategy. S72. Real-time monitoring of actual user response data of the currently deployed fine-grained marketing strategy within a predetermined period, and calculation of a strategy matching score; S73. Continuously calculate the strategy matching scores of the fine-grained marketing strategies, and compare the latest strategy matching scores with the set performance decay threshold in real time to determine whether the performance of the current marketing strategy is continuously below the performance decay threshold; S74. When the strategy matching score of a marketing strategy falls below the performance attenuation threshold for multiple consecutive times, the marketing strategy retirement mechanism is automatically activated to decommission the currently deployed fine-grained marketing strategy in real time. S75. When the current marketing strategy is retired, steps S3 to S6 are re-triggered to generate a new alternative fine-grained marketing strategy using the R3GAN strategy generation model that has undergone real-time self-perception adaptation and adjustment. S76. The strategy parameters of the newly generated alternative fine-grained marketing strategy are optimized and adjusted through the Feihu optimization algorithm with a double-layer nested structure, and finally deployed on the e-commerce platform to replace the retired marketing strategy.

Images