Marketing activity optimization promotion method and system based on generative model

Abstract

The invention relates to the technical field of marketing optimization, and discloses a marketing activity optimization promotion method and system based on a generative model. The method comprises the steps that current marketing activity content and real-time market data are collected, and related historical records are retrieved; constructing a market environment graph, and extracting a structured environment feature vector by using a graph neural network; calculating a historical marketing effectiveness score based on the features and historical effect data; a time sequence generation model is adopted to expand an efficiency sequence, and data abundance is improved; analyzing the sequence trend to obtain a market acceptance rate and a popularization effect attenuation rate; projecting the real-time environment data to a historical feature space, and correcting the parameters through similarity matching; and generating an optimization strategy by a strategy learning model in combination with the correction parameters. According to the method, the market complex association is deeply described through the graph neural network, and the data sparsity is overcome by using the generative model, so that the trend prediction is more accurate, and the strategy generation better fits the dynamic market environment.

Description

Technical Field

[0001] This invention relates to the field of marketing optimization technology, specifically to a method and system for optimizing and promoting marketing activities based on generative models. Background Technology

[0002] Existing marketing campaign optimization methods typically rely on statistical analysis of historical data. These methods deconstruct the market environment into a series of independent characteristic variables and establish statistical relationship models between these variables and marketing effectiveness. Such models treat the market as a static and divisible set of elements, ignoring the complex dynamic interaction networks between various entities within the market. The interrelationships between entities, such as word-of-mouth marketing, competitive effects, and channel synergy, are difficult to quantify and incorporate into the analytical framework, resulting in a superficial portrayal of the market environment that fails to reflect its inherent structural dynamics.

[0003] In trend forecasting, conventional techniques rely on limited historical marketing data points for curve fitting or time series analysis to infer key parameters such as market acceptance speed or effect decay cycles. However, historical marketing campaign data is often limited in quantity and contains significant noise, especially when facing new market scenarios or unexpected events, where the representativeness and coverage of historical data are particularly problematic. This data bottleneck results in poor stability and weak generalization ability for trend analysis based on purely historical data, making it difficult to effectively predict unseen but potentially occurring market dynamics, thus affecting the accuracy of subsequent strategy generation. Summary of the Invention

[0004] The purpose of this invention is to provide a marketing campaign optimization and promotion method and system based on a generative model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a marketing campaign optimization and promotion method based on a generative model, the method comprising: Collect the main content of the current marketing campaign and real-time market environment data, and retrieve historical marketing records associated with the main content of the marketing campaign; construct a market environment graph with market entities as nodes and relationships between entities as edges, and use a graph neural network to embed features into historical market environment information to obtain environmental feature vectors; Based on the environmental feature vector and the corresponding historical marketing performance data, a historical marketing effectiveness score is generated; a time series generation model is used to augment the historical marketing effectiveness score to form an expanded marketing effectiveness sequence. Trend analysis was performed on the expanded marketing effectiveness sequence to obtain market acceptance rate and promotion effect decay rate; Real-time market environment data is projected onto the historical environment feature space, a reference environment is selected through feature similarity calculation, and the market acceptance rate and promotion effect decay rate are corrected. By combining the corrected market acceptance rate and the promotion effect decay rate, a marketing promotion optimization strategy is generated through a strategy learning model.

[0006] Preferably, the main content of the current marketing campaign and real-time market environment data are collected, and historical marketing records associated with the main content of the marketing campaign are retrieved, including: Obtain a detailed description of the current marketing campaign, and simultaneously collect real-time user behavior data and competitive landscape data from the marketing campaign publishing platform as real-time market environment data; based on the keywords of the main content of the marketing campaign, perform semantic matching retrieval in the historical campaign database to obtain a set of related historical marketing records.

[0007] Preferably, a market environment graph is constructed with market entities as nodes and relationships between entities as edges. A graph neural network is then used to embed features from historical market environment information to obtain an environmental feature vector, including: Identify key market entities in historical market environment information and define the interaction types between entities as edge attributes to construct a dynamic market environment graph; initialize the number of layers and attention mechanism parameters of the graph neural network; aggregate node neighbor information through the message passing mechanism of the graph neural network, update node representations, and output the environmental feature vectors of all nodes.

[0008] Preferably, based on the environmental feature vector and the corresponding historical marketing performance data, a historical marketing effectiveness score is calculated and generated, including: Multiple key performance indicators (KPIs) are extracted from historical marketing performance data and standardized. The environmental feature vector and the standardized KPIs are then input into a fully connected network to calculate historical marketing effectiveness scores. The weight parameters of the fully connected network are optimized using the gradient descent algorithm to minimize the effectiveness score prediction error.

[0009] Preferably, a time series generation model is used to augment historical marketing effectiveness scores, forming an expanded marketing effectiveness series, including: The historical marketing effectiveness scores are arranged in chronological order to form the initial time series; the distribution characteristics of the initial time series are learned using a recurrent generative adversarial network to generate a synthetic effectiveness score; the synthetic effectiveness score is inserted into the gaps of the initial time series and smoothed to form the expanded marketing effectiveness series.

[0010] Preferably, trend analysis is performed on the extended marketing effectiveness sequence to obtain market acceptance rate and promotion effect decay rate, including: The temporal convolutional network model is used to extract local and global trend features of the expanded marketing effectiveness sequence; the trend features are input into the regression layer to output the initial estimates of market acceptance rate and promotion effect decay rate; the temporal convolutional network model is optimized through the validation set to obtain the final market acceptance rate and promotion effect decay rate.

[0011] Preferably, real-time market environment data is projected onto a historical environment feature space, a reference environment is selected through feature similarity calculation, and the market acceptance rate and promotion effect decay rate are corrected, including: The real-time market environment data is encoded into a real-time environment feature vector using an encoder; the cosine similarity between the real-time environment feature vector and the historical environment feature vector is calculated, and the historical environment with the highest similarity is selected as the reference environment; based on the Euclidean distance between the real-time environment feature vector and the reference environment feature vector, a correction factor is calculated to perform weighted correction on the market acceptance rate and the promotion effect decay rate.

[0012] Preferably, by combining the corrected market acceptance rate and the promotion effect decay rate, a marketing promotion optimization strategy is generated through a strategy learning model, including: Construct the state space of the strategy learning model, where the states include the corrected market acceptance rate and the promotion effect decay rate; define the action space as possible promotion strategy operations; use a deep Q-network to learn the state-action value function, and select the optimal promotion strategy through an epsilon-greedy strategy. The deep Q-network is used to learn the state-action value function, and the optimal promotion strategy is selected through an epsilon-greedy strategy, including: initializing the main network and target network of the deep Q-network; storing state transition samples through experience replay; and periodically updating the target network parameters to reduce correlation and stabilize the training process.

[0013] Preferably, a recurrent generative adversarial network is used to learn the distribution features of the initial time series to generate a synthetic efficacy score, including: The generator is trained to generate synthetic performance scores based on random noise and historical context; the discriminator is trained to distinguish between real and synthetic performance scores; and the authenticity of the synthetic data is improved through adversarial training between the generator and the discriminator.

[0014] Preferably, the present invention also includes a marketing campaign optimization and promotion system based on a generative model, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the marketing campaign optimization and promotion method based on the generative model described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By constructing a market environment graph with market entities as nodes and inter-entity relationships as edges, and utilizing graph neural networks for feature embedding, discrete market entities and their complex relationships are integrated into a low-dimensional, dense environmental feature vector. This method can capture structured information that transcends individual attributes. The representation of the market environment shifts from an isolated list of indicators to a deep encoding of the overall relationship network, making the understanding of the context of marketing activities more global and dynamic, and providing richer and more essential environmental contextual information for subsequent analysis.

[0016] By employing a time-series generative model to augment historical marketing effectiveness scores, the constraints of limited historical data samples are overcome. This model learns the inherent distribution and evolutionary patterns of historical effectiveness sequences and generates synthetic sequences that conform to real statistical laws but have not been actually observed in history. This process enriches the data foundation for trend analysis, ensuring that dynamic parameters such as market acceptance rate and promotional effect decay rate, extracted from the expanded sequences, not only rely on limited past experience but also incorporate the potential market patterns learned by the model. The resulting trend parameters possess stronger anti-overfitting capabilities and robustness in inferring unknown market scenarios.

[0017] Real-time environmental features encoded by a graph network are projected onto a historical feature space for similarity matching, and trend parameters derived from the generative model are corrected accordingly. This achieves adaptive calibration between static historical patterns and dynamic real-time environments. Based on the similarity between real-time environmental features and historical reference environments, the expected market acceptance and attenuation are dynamically adjusted, enabling the parameters upon which the strategy is based to respond sensitively to subtle changes in the market environment, thereby improving the timeliness and context relevance of the strategy. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the marketing campaign optimization and promotion method based on the generative model described in this invention. Figure 2 A flowchart for constructing a market environment graph and embedding graph neural network features; Figure 3 A flowchart for generating historical marketing effectiveness scores; Figure 4 This is a comparison chart of market acceptance trends based on TCN optimization. Figure 5 This is an annual trend chart showing the market acceptance rate and the rate of decline in promotional effectiveness. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a marketing campaign optimization and promotion method based on a generative model. The method includes: collecting the main content of the current marketing campaign and real-time market environment data, retrieving historical marketing records associated with the main content of the marketing campaign, constructing a market environment graph with market entities as nodes and relationships between entities as edges, embedding features into historical market environment information using a graph neural network to obtain an environment feature vector, generating a historical marketing effectiveness score based on the environment feature vector and corresponding historical marketing effect data, expanding the historical marketing effectiveness score using a time series generation model to form an expanded marketing effectiveness sequence, performing trend analysis on the expanded marketing effectiveness sequence to obtain the market acceptance rate and promotion effect decay rate, projecting real-time market environment data onto the historical environment feature space, selecting a reference environment through feature similarity calculation, and correcting the market acceptance rate and promotion effect decay rate, and generating a marketing promotion optimization strategy by combining the corrected market acceptance rate and promotion effect decay rate through a strategy learning model.

[0021] Example 1: See Figure 2In specific implementation, when collecting the main content of the current marketing campaign and real-time market environment data, a detailed description of the current marketing campaign is obtained. Simultaneously, real-time user behavior data and competitive landscape data from the marketing campaign's publishing platform are collected as real-time market environment data. Based on keywords in the main content of the marketing campaign, semantic matching retrieval is performed in the historical campaign database to obtain a set of related historical marketing records. When constructing a market environment graph with market entities as nodes and relationships between entities as edges, key market entities in the historical market environment information are identified, and the interaction types between entities are defined as edge attributes. A dynamic market environment graph is constructed, the number of layers and attention mechanism parameters of the graph neural network are initialized, and node neighbor information is aggregated through the message passing mechanism of the graph neural network to update node representations and output the environmental feature vectors of all nodes. In some embodiments, real-time user behavior data includes user click-through rate and dwell time, competitive landscape data includes competitor advertising frequency and market share, and semantic matching retrieval uses cosine similarity to calculate the matching degree between keyword vectors and historical record description vectors to obtain a set of highly relevant historical marketing records. It is understandable that key market entities include brand owners, consumer groups, and distribution channels. The types of interactions between these entities include transactional relationships and competitive relationships. The dynamic market environment graph represents the spatiotemporal changes of these entities and their relationships through a graph structure. In specific implementation, the message passing mechanism of the graph neural network updates node features through multi-layer propagation, where the node update formula is expressed as: , where: h v (l+1) Let W represent the feature vector of node v in layer l+1, σ represent the activation function, and W represent the feature vector of node v in layer l+1. (l) Let h represent the weight matrix of the l-th layer, AGGREGATE represent the aggregation function of the features of neighboring nodes, N(v) represent the set of neighboring nodes of node v, and h represent the weight matrix of the l-th layer. u (l) Let represent the feature vector of neighbor node u at layer l, and b(l) represent the bias term at layer l. Optionally, the attention mechanism parameters include the dimension settings of the query vector and key vector, used to weight the importance of neighbor nodes in message passing. In some embodiments, the number of layers of the initialized graph neural network is set to 3, the attention mechanism parameters are randomly initialized and adjusted during training, and the edge attributes of the dynamic market environment graph include relation strength and directionality. It is understood that the output dimension of the environment feature vector is fixed at 128 dimensions to be compatible with subsequent fully connected network processing.

[0022] Example 2: See Figure 3In specific implementation, when generating historical marketing effectiveness scores based on environmental feature vectors and corresponding historical marketing performance data, multiple key performance indicators (KPIs) are extracted from the historical marketing performance data, standardized, and then the environmental feature vectors and standardized KPIs are input into a fully connected network to calculate the historical marketing effectiveness scores. In some embodiments, the extracted KPIs include conversion rate, customer acquisition cost, and user engagement. The standardization process uses a min-max normalization method to map the values ​​of each indicator to the [0,1] interval. In specific implementation, the min-max normalization method involves extracting all numerical sequences of each KPI from the historical marketing performance data, calculating the minimum and maximum values ​​of conversion rate, customer acquisition cost, and user engagement in the entire historical dataset, and then performing a linear transformation on the original values ​​of each indicator. By subtracting the minimum value of the indicator and dividing by the difference between the maximum and minimum values, each data point is mapped to a closed interval between zero and one. This processing transforms indicator values ​​of different dimensions into a unified scale, eliminates the impact of numerical range differences on the training of the fully connected network, and improves the stability and consistency of historical marketing effectiveness score calculation. It can be understood that the environmental feature vector is a 128-dimensional vector, and the standardized key performance indicators (KPIs) form a 3-dimensional vector. The fully connected network concatenates these inputs into a 131-dimensional vector for subsequent processing. In specific implementation, the fully connected network contains three hidden layers with 64, 32, and 16 neurons respectively. The final output layer generates the historical marketing effectiveness score through a linear transformation. The calculation process of the historical marketing effectiveness score is expressed by the following formula: Where: S represents the historical marketing effectiveness score, f outputThe output layer activation function is represented by W1, W2, and W3, which represent the weight matrices of each layer in the fully connected network, respectively. b1 and b2 represent the bias vectors of the hidden layers, b3 represents the bias scalar of the output layer, σ represents the activation function of the hidden layers, h represents the environmental feature vector, k represents the standardized key performance indicator vector, and [h;k] represents the vector concatenation operation. Optionally, the ReLU function is used for the hidden layer activation function, and the Sigmoid function is used for the output layer activation function to constrain the historical marketing effectiveness score to the range of 0 to 1. The weight parameters of the fully connected network are optimized using the gradient descent algorithm to minimize the effectiveness score prediction error. The gradient descent algorithm uses the mean squared error as the loss function and calculates the gradient of the weight parameters through backpropagation. In some embodiments, the learning rate of the gradient descent algorithm is set to 0.001, the batch size is set to 32, and the training period is set to 100 rounds. It can be understood that the weight parameters are initialized using the Xavier initialization method, the bias term is initialized to a zero vector, and the Adam optimizer is used to update the weight parameters of the fully connected network during the optimization process. Optionally, the prediction error of the historical marketing effectiveness score is calculated by comparing the output value of the fully connected network with the actual observed historical marketing effectiveness indicators, which are derived from real records in the historical database.

[0023] Example 3: In a specific implementation, a time series generation model is used to augment historical marketing performance scores to form an expanded marketing performance sequence. The historical marketing performance scores are arranged chronologically as the initial time series. A recurrent generative adversarial network (RGAN) is used to learn the distribution characteristics of the initial time series to generate synthetic performance scores. These synthetic scores are then inserted into the gaps in the initial time series and smoothed to form the expanded marketing performance sequence. When generating synthetic performance scores, the generator is trained to generate synthetic performance scores based on random noise and historical context, while the discriminator is trained to distinguish between real and synthetic performance scores. Through adversarial training between the generator and the discriminator, the authenticity of the synthetic data is improved. In some embodiments, the initial time series is arranged in ascending order based on timestamps. The historical marketing performance scores come from the output of a fully connected network. The generator in the RGAN uses a long short-term memory (LSTM) network structure, and the discriminator uses a convolutional neural network structure. It can be understood that random noise is sampled from a standard normal distribution, historical context is extracted from the initial time series using a sliding window approach, the generator's long short-term memory network has a hidden layer dimension of 64, and the discriminator's convolutional neural network contains two convolutional layers and pooling layers. In specific implementations, the process of generating the synthetic performance score is expressed by the following formula: ,in: This represents the composite performance score generated at time step t. Represents a generator network, zt This represents a random noise vector at time step t. Represents a state vector based on historical context. This represents the set of parameters for the generator network. Optional, a random noise vector z. t The dimension is set to 16, and the historical context state vector is... The generator parameter set is obtained by encoding the performance score sequence of the previous time step. This includes the weight matrix and bias terms of the Long Short-Term Memory (LSTM) network. When training the discriminator to distinguish between real and synthetic performance scores, the discriminator outputs a probability value representing the likelihood that the input sequence is real data. The discriminator's loss function uses binary cross-entropy to calculate the difference between the predicted value and the true label. In some embodiments, during adversarial training, the generator and discriminator alternately update parameters. The generator's optimization objective is to maximize the discriminator's misclassification probability of synthetic data, while the discriminator's optimization objective is to accurately distinguish between real and synthetic performance scores. It can be understood that when synthetic performance scores are inserted into the gaps of the initial time series, the insertion position is selected based on time order and sequence continuity, and smoothing uses a moving average method to eliminate discontinuities at the insertion point. Optionally, the moving average window size is set to 5 time steps, and the length of the expanded marketing performance sequence is adjusted to twice that of the original sequence to provide richer data for subsequent trend analysis.

[0024] Example 4: In specific implementation, when performing trend analysis on the expanded marketing effectiveness sequence to obtain the market acceptance rate and promotion effect decay rate, a temporal convolutional network model is applied to extract the local and global trend features of the expanded marketing effectiveness sequence. The trend features are input into the regression layer, and the initial estimates of the market acceptance rate and promotion effect decay rate are output. The temporal convolutional network model is then fine-tuned using a validation set to obtain the final market acceptance rate and promotion effect decay rate. Real-time market environment data is projected onto the historical environment feature space, and a reference environment is selected through feature similarity calculation. When correcting the market acceptance rate and promotion effect decay rate, an encoder is used to encode the real-time market environment data into a real-time environment feature vector. The cosine similarity between the real-time environment feature vector and the historical environment feature vector is calculated, and the historical environment with the highest similarity is selected as the reference environment. Based on the Euclidean distance between the real-time environment feature vector and the reference environment feature vector, a correction factor is calculated to perform weighted correction on the market acceptance rate and promotion effect decay rate. In some embodiments, the kernel size of the temporal convolutional network model is set to 3 and 5 to capture trend features at different time scales. The regression layer adopts a fully connected layer structure, and the outputs two scalar values ​​corresponding to market acceptance rate and promotion effect decay rate, respectively. It can be understood that the expanded marketing effectiveness sequence comes from the output of the recurrent generative adversarial network, the validation set contains some historical data for model tuning, and the encoder is a pre-trained neural network model. In a specific implementation, the correction factor is calculated using the following formula: Where: γ represents the correction factor, λ represents the distance scaling parameter, and d(v real ,v ref ) represents the real-time environment feature vector v real With reference environment feature vector v ref The Euclidean distance between them. Optionally, the distance scaling parameter λ is set to 0.1, the Euclidean distance is calculated using the L2 norm, and weighted correction is achieved by multiplying the market acceptance rate by a correction factor γ and the promotion effect decay rate by a correction factor γ. In some embodiments, cosine similarity is calculated using the ratio of the vector dot product to the modulus, selecting the top k historical environments with the highest similarity as candidates, and then refining the selection based on the Euclidean distance. It can be understood that both the real-time environment feature vector and the historical environment feature vector are 128-dimensional vectors, and the dimension of the encoder output vector is consistent with that of the historical environment feature vector. Refer to Table 1 for the calculation of feature similarity.

[0025] Table 1: Feature Similarity Calculation Table

[0026] Optionally, the feature vectors in the table display the first three dimensions, the cosine similarity is rounded to two decimal places, and the reference environment is selected based on the highest similarity value. It can be understood that the weighted adjusted market acceptance rate and promotion effect decay rate are used as input to the subsequent strategy learning model.

[0027] See Figure 4 In generative model-based marketing campaign optimization methods, the trend analysis of market acceptance relies on Temporal Convolutional Network (TCN) tuning techniques. Specifically, the initial estimated market acceptance rate is a discrete, fluctuating sequence that changes over time. After TCN tuning, a smooth market acceptance rate sequence is formed. The difference between the two reflects the effectiveness of TCN in extracting and tuning the local and global trend features of the marketing effectiveness sequence: the initial estimated sequence fluctuates wildly, reflecting the unstable estimation of market acceptance rate without model tuning; the sequence after TCN tuning shows a steady upward trend, indicating that after extracting features through the temporal convolutional network model and tuning through a validation set, a more reliable estimate of market acceptance rate is obtained.

[0028] Example 5: In specific implementation, when generating marketing promotion optimization strategies through a strategy learning model by combining the corrected market acceptance rate and promotion effect decay rate, the state space of the strategy learning model is constructed. The states include the corrected market acceptance rate and promotion effect decay rate. The action space is defined as possible promotion strategy operations. A deep Q-network is used to learn the state-action value function. The optimal promotion strategy is selected through an epsilon-greedy strategy. When using the deep Q-network to learn the state-action value function and selecting the optimal promotion strategy through an epsilon-greedy strategy, the main network and target network of the deep Q-network are initialized. State transition samples are stored through experience replay, and the parameters of the target network are updated periodically to reduce correlation and stabilize the training process. In some embodiments, the state space is a two-dimensional continuous space, with the two dimensions corresponding to the corrected market acceptance rate and promotion effect decay rate, respectively. The action space includes three basic operations: increasing advertising placement, adjusting the target audience, and modifying marketing content. It can be understood that the main network and target network of the deep Q-network have the same network structure, containing two fully connected hidden layers with 64 and 32 neurons per layer, respectively. The capacity of the experience replay buffer is set to 10,000 state transition samples. In practical implementation, the target Q value is calculated using the following formula: , where: Y target Let Q represent the target Q value, R represent the immediate reward, and ζ represent the discount factor. This indicates the use of the target network parameters in the next state S′. The maximum Q-value is calculated, and a′ represents the possible action in the next state S′. Optionally, the immediate reward R is calculated based on marketing performance metrics after strategy execution, the discount factor ζ is set to 0.99, and the target network parameters are... Updated after every 100 training iterations. In some embodiments, the initial exploration probability of the epsilon-greedy policy is set to 1.0, which linearly decays to 0.1 during training. State transition samples are stored in tuple form, containing four elements: current state, action performed, reward obtained, and next state. It can be understood that the loss function of the deep Q-network uses mean squared error to calculate the difference between the current Q-value and the target Q-value. The optimization process uses stochastic gradient descent to update the main network parameters, and experience replay randomly samples batch data from the buffer for training. In specific implementations, the stochastic gradient descent algorithm includes randomly sampling batch data from the experience replay buffer, calculating the gradient of the deep Q-network loss function with respect to the weight parameters, using mean squared error to calculate the difference between the current Q-value and the target Q-value, calculating the gradient through backpropagation, and then updating the weight parameters of the main network using a preset learning rate. This process is iterative to minimize the loss function. In some embodiments, the learning rate is set to 0.001, the batch size is set to 32, the training period is set according to the model convergence, and the optimization process gradually adjusts the network parameters through multiple iterations to ensure training stability and policy learning effectiveness. Optionally, the specific parameters for promotion strategy operations are determined through an action mapping table. Increasing the advertising placement operation corresponds to increasing the budget allocation ratio, adjusting the target audience operation corresponds to modifying the user profile filtering conditions, and modifying the marketing content operation corresponds to updating the creative material version.

[0029] See Figure 5 The chart presents the dynamic changes in market acceptance and promotional effectiveness decay rate throughout the year. Market acceptance, represented by blue bars, shows a continuous upward trend, gradually increasing from 0.61 in January to 0.89 in December, reflecting the increasing market acceptance of the marketing campaign over time. Promotional effectiveness decay rate, represented by purple line graphs, is relatively flat in the early stages (January-May), then rises significantly from June, peaking at 0.137 in September before fluctuating slightly, reaching 0.136 in December. This trend indicates that as market acceptance increases, the rate of promotional effectiveness decay accelerates in the middle stage before stabilizing, demonstrating the dynamic relationship between market acceptance and promotional effectiveness decay. This provides data support for optimizing marketing strategies, allowing for adjustments to promotional intensity and methods at different stages to balance the relationship between market acceptance and promotional effectiveness decay.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A marketing campaign optimization and promotion method based on a generative model, characterized in that, The method includes: Collect the main content of the current marketing campaign and real-time market environment data, and retrieve historical marketing records associated with the main content of the marketing campaign; construct a market environment graph with market entities as nodes and relationships between entities as edges, and use a graph neural network to embed features into historical market environment information to obtain environmental feature vectors; Based on the environmental feature vector and the corresponding historical marketing performance data, a historical marketing effectiveness score is generated; a time series generation model is used to augment the historical marketing effectiveness score to form an expanded marketing effectiveness sequence. Trend analysis was performed on the expanded marketing effectiveness sequence to obtain market acceptance rate and promotion effect decay rate; Real-time market environment data is projected onto the historical environment feature space, a reference environment is selected through feature similarity calculation, and the market acceptance rate and promotion effect decay rate are corrected. By combining the corrected market acceptance rate and the promotion effect decay rate, a marketing promotion optimization strategy is generated through a strategy learning model.

2. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, Collect the main content of the current marketing campaign and real-time market environment data, and retrieve historical marketing records associated with the main content of the marketing campaign, including: Obtain a detailed description of the current marketing campaign, and simultaneously collect real-time user behavior data and competitive landscape data from the marketing campaign publishing platform as real-time market environment data; based on the keywords of the main content of the marketing campaign, perform semantic matching retrieval in the historical campaign database to obtain a set of related historical marketing records.

3. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, A market environment graph is constructed with market entities as nodes and relationships between entities as edges. A graph neural network is then used to embed historical market environment information to obtain an environmental feature vector, including: Identify key market entities in historical market environment information and define the interaction types between entities as edge attributes to construct a dynamic market environment graph; initialize the number of layers and attention mechanism parameters of the graph neural network; aggregate node neighbor information through the message passing mechanism of the graph neural network, update node representations, and output the environmental feature vectors of all nodes.

4. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, Based on the environmental feature vector and the corresponding historical marketing performance data, a historical marketing effectiveness score is calculated and generated, including: Multiple key performance indicators (KPIs) are extracted from historical marketing performance data and standardized. The environmental feature vector and the standardized KPIs are then input into a fully connected network to calculate historical marketing effectiveness scores. The weight parameters of the fully connected network are optimized using the gradient descent algorithm to minimize the effectiveness score prediction error.

5. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, A time series generation model is used to augment historical marketing effectiveness scores, resulting in an expanded marketing effectiveness series, including: The historical marketing effectiveness scores are arranged in chronological order to form the initial time series; the distribution characteristics of the initial time series are learned using a recurrent generative adversarial network to generate a synthetic effectiveness score; the synthetic effectiveness score is inserted into the gaps of the initial time series and smoothed to form the expanded marketing effectiveness series.

6. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, Trend analysis was performed on the extended marketing effectiveness sequence to obtain market acceptance rate and promotion effect decay rate, including: The temporal convolutional network model is used to extract local and global trend features of the expanded marketing effectiveness sequence; the trend features are input into the regression layer to output the initial estimates of market acceptance rate and promotion effect decay rate; the temporal convolutional network model is optimized through the validation set to obtain the final market acceptance rate and promotion effect decay rate.

7. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, Projecting real-time market environment data onto a historical environment feature space, selecting a reference environment through feature similarity calculation, and correcting for the market acceptance rate and promotion effect decay rate, including: The real-time market environment data is encoded into a real-time environment feature vector using an encoder; the cosine similarity between the real-time environment feature vector and the historical environment feature vector is calculated, and the historical environment with the highest similarity is selected as the reference environment; based on the Euclidean distance between the real-time environment feature vector and the reference environment feature vector, a correction factor is calculated to perform weighted correction on the market acceptance rate and the promotion effect decay rate.

8. The marketing campaign optimization and promotion method based on a generative model according to claim 1, characterized in that, By combining the corrected market acceptance rate and promotional effectiveness decay rate, a marketing promotion optimization strategy is generated through a strategy learning model, including: Construct the state space of the strategy learning model, where the states include the corrected market acceptance rate and the promotion effect decay rate; define the action space as possible promotion strategy operations; use a deep Q-network to learn the state-action value function, and select the optimal promotion strategy through an epsilon-greedy strategy. The deep Q-network is used to learn the state-action value function, and the optimal promotion strategy is selected through an epsilon-greedy strategy, including: initializing the main network and target network of the deep Q-network; storing state transition samples through experience replay; and periodically updating the target network parameters to reduce correlation and stabilize the training process.

9. The marketing campaign optimization and promotion method based on a generative model according to claim 5, characterized in that, The distribution characteristics of the initial time series are learned using a recurrent generative adversarial network to generate a synthetic efficacy score, including: The generator is trained to generate synthetic performance scores based on random noise and historical context; the discriminator is trained to distinguish between real and synthetic performance scores; and the authenticity of the synthetic data is improved through adversarial training between the generator and the discriminator.

10. A marketing campaign optimization and promotion system based on a generative model, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the marketing campaign optimization and promotion method based on the generative model as described in any one of claims 1 to 9.

Images