Marketing strategy recommendation model training method based on big data

Abstract

The invention relates to the technical field of big data, and discloses a marketing strategy recommendation model training method based on big data, comprising the following steps: S1, collecting data; s2, performing sample generation by adopting a generative adversarial network algorithm based on sparse learning to realize data expansion and overall enhancement of a data set; s3, inputting the expanded data into a pre-trained classifier model to obtain a classification result; the classifier model adopts an extreme learning machine algorithm based on fractional differential as a classification algorithm; and S4, determining a marketing strategy result according to a classification result. The method has the beneficial effects that the non-uniform features in the commodity information are weighted through a sparse learning mechanism, key information is emphasized, redundant features are inhibited, and the training efficiency and classification precision of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of big data technology, specifically a method for training a marketing strategy recommendation model based on big data. Background Technology

[0002] With the widespread adoption of the internet and e-commerce platforms, the quantity and complexity of product information have increased dramatically. Businesses face the challenge of extracting valuable information from massive amounts of data and developing effective marketing strategies. Traditional marketing strategy development methods rely primarily on human experience and simple data analysis, which are ill-suited to the diversity and dynamic changes of modern product information. Furthermore, product information data typically exhibits high dimensionality, non-linearity, and sparsity, making traditional machine learning models ineffective in processing this data, resulting in suboptimal classification accuracy and recommendation performance.

[0003] Based on this, the existing technology has the following problems: 1. In existing technologies, the collection and labeling of product information data is time-consuming and labor-intensive, resulting in a limited number of training samples, insufficient model generalization ability, and a tendency to overfit. Traditional data augmentation methods cannot effectively handle the non-uniform distribution and complex feature relationships of product information data.

[0004] 2. Traditional extreme learning machines or other classification algorithms may not be able to effectively capture non-linear features when processing complex product information data, resulting in low classification accuracy.

[0005] 3. There are a lot of redundant features in the product information data. Existing technologies lack effective sparsity processing mechanisms, resulting in low model training efficiency and susceptibility to noise and outliers.

[0006] 4. Existing model parameter update mechanisms are relatively fixed and cannot be adaptively adjusted according to dynamic changes in input data, making it difficult to cope with complex and ever-changing marketing scenarios. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for training a marketing strategy recommendation model based on big data.

[0008] To achieve the above objectives, the present invention employs the following technical solution: According to one aspect of the present invention, a method for training a marketing strategy recommendation model based on big data is provided, comprising the following steps: Collect data from the product information management system, product manuals, and product detail pages on e-commerce platforms; A generative adversarial network algorithm based on sparse learning is used to generate samples, which expands the collected data and enhances the overall dataset. The augmented data is input into a pre-trained classifier model to obtain the classification result; the classifier model uses the extreme learning machine algorithm based on fractional derivative as the classification algorithm; Based on the classification results, determine the marketing strategy outcomes.

[0009] Furthermore, the training method for generative adversarial networks based on sparse learning is as follows: Initialize the generator and discriminator of the generative adversarial network. The weights of the generator and the discriminator are initialized with random numbers drawn from a Gaussian distribution. During each iteration of training, the generator extracts non-uniform feature information from the input noisy data and assigns dynamic weights to the feature information. The non-uniform features are weighted through a sparse learning-based mechanism. The generator generates data based on the input noise data and the weighted feature information; After receiving samples generated by the generator and samples from real data, the discriminator distinguishes them through a deep network layer and finally outputs the judgment result. A diversity constraint term is used to optimize the diversity of the generated samples. The diversity constraint term is calculated based on the information entropy of the feature distribution of the generated samples. During each round of training, the generator and discriminator are subjected to adversarial training. In each iteration of training, the generator optimizes the weights based on the sparsity of the features, gradually reducing the interference of redundant features; Update the weights of the generator and discriminator.

[0010] Furthermore, the training method for the extreme learning machine algorithm based on fractional derivatives is as follows: The training data is input into the Extreme Learning Machine classifier, and the training data is mapped to a high-dimensional feature space through a nonlinear transformation. The weight matrix and bias terms of the extreme learning machine are initialized using the covariance matrix; In the hidden layers of the Extreme Learning Machine, the data is processed using a fractional-order differential strategy; Based on a fractional-order differential strategy, a self-similar curve-driven parameter learning method is used to update the weights of the extreme learning machine: The classification layer of the Extreme Learning Machine is the last layer, and the model output is converted into the predicted probability of each category through the Softmax function.

[0011] Furthermore, it also includes preprocessing the collected data, including: Data cleaning removes meaningless symbols, stop words, and duplicate information, while also performing word segmentation on the text. The cleaned text data is vectorized using the term frequency-inverse term frequency method. The vectors of product attributes, descriptions, and functions are concatenated to form a comprehensive feature vector; The concatenated feature vectors are normalized, and the normalized feature vectors are used as input data for the machine learning model.

[0012] Furthermore, each data record contains the product's attributes, product description, and product functions.

[0013] Furthermore, based on the sparse learning mechanism, the generator weights the non-uniform features, and the weighted features are: ; In the formula, These are dynamic learning parameters that represent the sparsity weights of features. For example, when generating a "promotional strategy", the weights of price-sensitive features are automatically increased. It is a non-uniform characteristic; It is a sparsification function that forces some unimportant features to be compressed to zero; This represents the non-uniform characteristics after weighting.

[0014] Furthermore, the fractional derivative not only relies on traditional gradient information but also considers information from historical states, and its calculation method is expressed as follows: ; In the formula, Denotes the fractional derivative. It is the order of the fractional differential; Let be the weight matrix of the extreme learning machine; This is a fractional weighting factor that adjusts the sensitivity of the fractional derivative to the current gradient. Let the loss function of the extreme learning machine be the gradient with respect to the weights. yes abbreviation, Indicates the first The loss function of the extreme learning machine in the next iteration with respect to the weight gradient, Let be the loss function of the extreme learning machine; Fractional dynamics representing time; This represents the integral operation, used to indicate the impact of past gradients on the current update; It is the step size of the historical iterations. It is different from The step size of the historical iterations; express differential.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a training method for a marketing strategy recommendation model based on big data. It employs a sparse learning mechanism to weight non-uniform features in product information, emphasizing key information and suppressing redundant features. A diversity constraint term optimizes the diversity of generated samples, preventing them from being overly concentrated on certain feature values. A dynamic weight adjustment strategy allows the generator and discriminator to gradually adapt to the complex distribution of product information data during training. Fractional derivatives can handle non-integer derivatives, capturing local nonlinear features and historical state information in the data, enhancing the model's adaptability to complex dynamic changes. Dynamic adjustment of fractional derivative parameters and the learning rate, along with a parameter learning method driven by self-similar curves to optimize the weight update mechanism, improves the model's training efficiency and classification accuracy. Attached Figure Description

[0016] Appendix Figure 1 This is a flowchart of the present invention; Appendix Figure 2 This is a diagram of the platform architecture of this invention. Detailed Implementation

[0017] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0018] like Figure 1 As shown, a method for training a marketing strategy recommendation model based on big data is disclosed, including... S1. Collect data from the product information management system, product manuals, and product detail pages on e-commerce platforms; S2. A generative adversarial network algorithm based on sparse learning is used to generate samples, expand the collected data, and enhance the overall dataset. S3. Input the expanded data into the pre-trained classifier model to obtain the classification result; the classifier model uses the extreme learning machine algorithm based on fractional derivative as the classification algorithm; S4. Based on the classification results, determine the marketing strategy results.

[0019] Specifically, the training data of this invention comes from the commodity information management system, product manuals, and product detail pages of e-commerce platforms. The data storage format adopts a structured CSV format, in which each data record contains commodity attributes (such as brand, model, material, etc.), commodity introduction (such as product features, usage scenarios, etc.), and commodity functions (such as function description, technical parameters, etc.). The data labeling categories include: promotional strategies, new product promotion strategies, high-end product strategies, etc.

[0020] In order to transform the collected text data into a discrete format suitable for machine learning model input, this invention first preprocesses the collected text data, specifically including: S101. Remove meaningless symbols, stop words, and duplicate information, and perform word segmentation on the text to ensure the cleanliness and consistency of the text data. S102. The cleaned text data is vectorized using the Term Frequency-Inverse Document Frequency (TF-IDF) method. The TF-IDF method calculates the weight of each word for the product's attributes, description, and function, and converts them into high-dimensional vectors to effectively reflect the features of the text content, while avoiding feature differences caused by varying text lengths. S103. Concatenate the vectors of product attributes, descriptions, and functions to form a comprehensive feature vector; S104. Normalize the concatenated feature vectors to ensure that features of different dimensions are comparable. The normalized feature vectors are used as input data for the machine learning model.

[0021] The data augmentation module augments the collected data by expanding the number of samples through a data augmentation model, thereby enhancing the training dataset.

[0022] It is understandable that in the task of this invention, the acquisition, annotation, and preprocessing of training data are time-consuming and labor-intensive, and insufficient training samples can easily lead to poor model generalization ability and affect model accuracy. This invention uses a generative adversarial network algorithm based on sparse learning to generate samples, thereby achieving data augmentation and overall enhancement of the dataset.

[0023] Generative Adversarial Network (GAN) algorithms consist of two components: a generator and a discriminator. The generator aims to transform the input noisy data to generate synthetic samples that are closer to the real training data. The discriminator is responsible for distinguishing between the generated samples and the real samples, forcing the generator to generate more realistic data.

[0024] Unlike traditional generative adversarial networks (GANs), this invention employs a sparsity-based generation strategy during the training process. Product information data typically includes attributes such as brand, model, and material. These attributes may exhibit a non-uniform distribution in the data, meaning that some attribute values ​​may appear frequently while others appear less frequently. Specifically, the training process of the sparse learning-based GAN algorithm is as follows: S201. Initialize the generator and discriminator of the generative adversarial network. The weights of the generator are initialized with random numbers drawn from a Gaussian distribution, expressed as: ; In the formula, To conform to a specific distribution; The weights of the generator, The initial variance of the generator weights determines the initial distribution range of the generated samples; It follows a normal distribution.

[0025] Furthermore, the weights of the discriminator are initialized with random numbers drawn from a Gaussian distribution, expressed as: ; In the formula, The weights of the discriminator.

[0026] S202. In each iteration of training, the generator extracts non-uniform feature information from the input noisy data and assigns dynamic weights to these features. The non-uniform features are weighted using a sparse learning-based mechanism to emphasize key information in the product data during generation. Specifically, the non-uniform features extracted by the generator are obtained from the noise through a multilayer perceptron, and are represented as follows: ; In the formula, It is a non-uniform characteristic; Represents a multilayer perceptron model; It is the input noise of the generator.

[0027] Furthermore, based on the sparse learning mechanism, the generator weights the non-uniform features, and the weighted features are: ; In the formula, These are dynamic learning parameters that represent the sparse weights of the features; It is a sparsification function that forces some unimportant features to be compressed to zero; This represents the non-uniform characteristics after weighting.

[0028] S203. The generator generates data based on the input noisy data and weighted feature information. The generator takes both non-uniform features and noisy data as input and generates new sample data through layer-by-layer transformations. In each training round, the generator gradually learns to generate samples that are closer and closer to the real data distribution through continuous parameter updates, as shown below: ; In the formula, These are transformation functions in the generator (e.g., convolution operations and fully connected layers in a convolutional neural network). These are samples generated by the generator.

[0029] In one embodiment, the process of generating samples is implemented through a deep neural network, which consists of multiple layers of nonlinear transformations. Through hierarchical mapping, the generator can generate the final samples based on input features and noise, as shown below: ; In the formula, and These are two activation functions (such as ReLU, LeakyReLU, or Sigmoid). This is a function for a deep neural network model.

[0030] S204. After receiving samples generated by the generator and samples from real data, the discriminator distinguishes them through a deep network layer and finally outputs the judgment result. The discriminator not only determines whether a sample is a real sample, but also scores the accuracy of its features. The discriminator's goal is to maximize the difference between real samples and generated samples, while minimizing misclassification of generated samples, as expressed as: ; In the formula, Indicates the discriminator; It is a deep convolutional network for the discriminator; The samples input to the discriminator may be samples generated by the generator or samples of real data; The weights of the discriminator.

[0031] In one embodiment, the discriminator is a deep convolutional network, consisting of multiple layers of convolution and pooling operations. The output of each layer is processed by a non-linear activation function, expressed as: ; In the formula, It is the activation function of the nth layer of the discriminator. It is the activation function of the first layer of the discriminator. It is the activation function of the (n-1)th layer of the discriminator; It is a max pooling operation used to reduce the dimensionality of features.

[0032] S205. The diversity of product information data is reflected in multiple aspects, such as usage scenarios and product features in product descriptions. The values ​​of these attributes may have a large range of variation. Therefore, diversity constraints are needed to ensure the diversity of generated samples. This invention uses diversity constraints to optimize the diversity of samples generated by the generator. The diversity constraints are calculated based on the information entropy of the feature distribution of the generated samples. The generator not only needs to obtain a high score in the discriminator, but also needs to ensure that the distribution of generated samples in the feature space is more uniform, thereby improving the diversity of generated samples. Specifically, the discretized information entropy of the feature distribution of generated samples is defined as follows: ; In the formula, Discretized information entropy for generating the feature distribution of samples; The first sample generated by the generator One eigenvalue; It is a feature The probability distribution of a feature value is the probability of that feature value appearing in the generated samples.

[0033] Furthermore, to ensure the diversity of generated samples, we want the information entropy to be as high as possible. This means that the feature distribution of generated samples needs to maintain a relatively large degree of uncertainty to avoid excessive concentration of certain feature values. Therefore, the feature distribution of generated samples is estimated by calculating the histogram of generated samples. Specifically, the feature values ​​of the samples are discretized into... If there are several intervals, the probability distribution is estimated by the frequency of the sample in each interval, and the calculation method is expressed as follows: ; In the formula, It is the feature value in the sample Number of times it appears This represents the total number of samples generated.

[0034] S206. During each training round, the generator and discriminator undergo adversarial training. The generator's loss function is calculated as follows: ; In the formula, The loss function for the generator; It is a hyperparameter that balances adversarial loss and diversity loss, controlling the impact of the diversity constraint term on the total loss; This indicates that the input noise is The expectation of the generator; This indicates the discriminator's judgment result on the generated sample; Represents the L1 norm; for The L1 norm represents the sparsity of the features. To control the hyperparameters of sparsity intensity.

[0035] Furthermore, the discriminator aims to maximize its ability to distinguish between real and generated samples. The discriminator's loss function is calculated as follows: ; In the formula, The loss function of the discriminator; This indicates that the input is real data. The expectation of the discriminator; This indicates that the generator generates data based on the input. The expectation of the discriminator; This represents the discriminator's judgment result on the real sample. This indicates the discriminator's judgment result on the generated sample.

[0036] S207. Product information data may contain a large number of redundant features. For example, some attribute values ​​may not have any practical meaning in describing product information. Sparse learning mechanisms can identify and suppress these redundant features, thereby improving the quality of generated samples. In each iteration of training, the generator optimizes the weights according to the sparsity of the features, gradually reducing the interference of redundant features. By adopting dynamic weight adjustment based on sparse learning mechanisms, the generator can focus more on generating samples with key features. The discriminator also adapts to these sparse features in the same process, gradually strengthening its ability to distinguish important features. The sparsity function is implemented by adding a regularization term to encourage some unimportant parts of the feature vector to tend to zero. Let... From input noise Extracted feature vectors, From input noise The extracted i-th feature vector; From input noise The extracted nth feature vector; the sparsity function is calculated as follows: ; In the formula, As a sparsity control factor, It is the Softmax function. Represents element-wise multiplication; It is a threshold function.

[0037] Furthermore, the threshold function compresses certain elements in the feature vector to zero, and its calculation method is expressed as follows: ; In the formula, It is a set threshold; only when the absolute value of the feature value is greater than 10 ... Only retain its original value if it is true; otherwise, set it to zero.

[0038] S208. As the training process continues, the number of samples output by the generator will gradually increase, and the quality of the samples will continuously improve. The weights of both the generator and the discriminator will be updated, and the update method is expressed as follows: ; Furthermore, the discriminator update process is represented as follows: ; In the formula, The learning rate for generating adversarial networks; For parameter update operations; The symbol represents the partial derivative.

[0039] S209. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000.

[0040] The expanded training data is then input into the classifier model of the marketing strategy recommendation module to perform marketing strategy recommendations.

[0041] This invention employs a fractional derivative-based Extreme Learning Machine (ELM) algorithm as the classification algorithm. Fractional derivatives enable the ELM model to be optimized in the non-integer domain, effectively capturing the complex dynamic changes of the input data, while increasing the model's flexibility and nonlinear characteristics. Compared with traditional ELM algorithms, the fractional derivative-based ELM algorithm proposed in this invention further improves classification accuracy by dynamically adjusting the fractional derivative parameters to cope with the distribution characteristics of different input data.

[0042] Specifically, the training process for the Limit Learning Machine algorithm based on fractional derivatives is as follows: S301. Product information data exhibits complex nonlinear relationships. For example, different brands and models of products may have nonlinear differences in function and usage scenarios. Extreme Learning Machines (ELMs) use nonlinear transformations to map input feature vectors to a high-dimensional feature space, enhancing the nonlinear expressive power of the data to capture the complex relationships between product features. This allows the model to classify more accurately. Specifically, training data is input into the ELM classifier, and the training data is mapped to a high-dimensional feature space through a nonlinear transformation, as shown below: ; In the formula, The feature vectors are input to the extreme learning machine; Here is the weight matrix of the extreme learning machine. This is the bias term for the Extreme Learning Machine; Use the Sigmoid activation function; This represents the high-dimensional feature representation after the hidden layer mapping of the Extreme Learning Machine.

[0043] S402. The weight matrix initialization of the Extreme Learning Machine incorporates statistical characteristics of the data, such as the mean, variance, and covariance matrix. This adapts to the data distribution and includes correlation information between input features. For example, for the correlation between different attributes in product information data (such as the association between brand and function), the calculation of the covariance matrix can better reflect this relationship, thus providing more reasonable initial weights for model training. The initialization method is expressed as follows: ; In the formula, is the regularization coefficient of the extreme learning machine, used to control the size of the weights; Represents a diagonal matrix; A standard deviation vector representing the characteristics of the data; The covariance matrix of the input features; It is a tiny constant used to avoid singular matrices; It is an identity matrix to ensure the stability of the matrix.

[0044] Furthermore, the initialization method of the bias term takes into account the mean vector of the data, which can adjust the initial value of the bias term according to the overall distribution of the data. This helps the model better adapt to the data distribution in the early stages of training, improving the model's convergence speed and classification performance, as expressed below: ; In the formula, It is the bias term adjustment coefficient of the extreme learning machine; It is the mean vector. This is a small constant used to avoid singular matrices. Preferably, Set to 0.0001.

[0045] S403. Fractional differentiation can handle non-integer derivatives. When faced with training data exhibiting irregular changes, it can capture local nonlinear features in the data. This is significant for nonlinear variations that may exist in product information data (such as performance differences between products made of different materials). Fractional differentiation not only relies on traditional gradient information but also considers historical state information, enabling it to more comprehensively reflect the dynamic changes in the data, thereby enhancing the model's expressive power. In the hidden layers of the Extreme Learning Machine (ELM), the data is processed using a fractional differentiation strategy. Fractional differentiation can handle non-integer derivatives and capture local nonlinear features in training data exhibiting irregular changes. It also achieves adaptive adjustment during training, thereby enhancing the model's expressive power. During the training process of the ELM, the optimization objective of fractional differentiation is to minimize the loss function of the ELM, expressed as: ; In the formula, Denotes the fractional derivative. It is the order of the fractional differential; Let the loss function of the extreme learning machine be the gradient with respect to the weights. Let be the loss function of the Extreme Learning Machine.

[0046] Furthermore, the fractional derivative not only relies on traditional gradient information but also considers information from historical states, and its calculation method is expressed as follows: ; In the formula, Denotes the fractional derivative. It is the order of the fractional differential; Let be the weight matrix of the extreme learning machine; This is a fractional weighting factor that adjusts the sensitivity of the fractional derivative to the current gradient. Let the loss function of the extreme learning machine be the gradient with respect to the weights. yes abbreviation, Indicates the first The loss function of the extreme learning machine in the next iteration with respect to the weight gradient, Let be the loss function of the extreme learning machine; Fractional dynamics representing time; This represents the integral operation, used to indicate the impact of past gradients on the current update; It is the step size of the historical iterations. It is different from The step size of the historical iterations; express differential.

[0047] S404. Based on the fractional derivative strategy, the classifier weights are dynamically optimized in each training process. The algorithm dynamically adjusts the derivative order, learning rate, and weight update mechanism according to the changes in input data. For example, for dynamic changes that may occur in product information data (such as the adjustment of new product promotion strategies), the self-similar curve driven parameter learning method can adaptively adjust the model parameters by analyzing the similarity patterns of input data, thereby improving the training efficiency and accuracy of the model. To achieve efficient weight updates, this invention employs a self-similar curve-driven parameter learning method. By analyzing the similarity patterns of the input data, it adaptively adjusts the model parameters, thereby improving the training efficiency and accuracy of the model. The weight update method of the Extreme Learning Machine is expressed as follows: ; In the formula, For the updated weights of the Extreme Learning Machine, The learning rate of the extreme learning machine. Adjusting parameters for the self-similarity of the Extreme Learning Machine; This is a fractional weighting factor that adjusts the sensitivity of the fractional derivative to the current gradient. This is an adaptive adjustment factor; This represents the current iteration number. Preferably, Set to 0.01, Set to 0.3, Set to 0.5. Set to 1.

[0048] S405, the classification layer of the Extreme Learning Machine is the last layer. The Softmax function converts the model output into predicted probabilities for each category, as follows: ; In the formula, For the Softmax function; The output of the classifier represents the predicted probability for each class; This is the Softmax activation function.

[0049] Furthermore, based on the output of the Extreme Learning Machine (ELM), its loss function is calculated. The model output is converted into predicted probabilities for each category using the Softmax function, and the loss function is then calculated. The model can accurately classify products into promotional strategies, new product launch strategies, or high-end product strategies based on their characteristics and labeled categories, thus achieving efficient product information data classification. The calculation method for the ELM loss function is expressed as follows: ; In the formula, This is a shorthand for the loss function of Extreme Learning Machine. For the number of categories, The first output of the classifier Class probability, For the first The true label of the class, It is an L2 norm; represents the regularization parameters for the Extreme Learning Machine.

[0050] S406. Repeat the above steps until a preset stopping iteration condition is met, indicating that the model training is complete. In one embodiment, the preset stopping iteration condition is reaching a preset maximum number of iterations. Preferably, the preset maximum number of iterations is set to 1000. Example: Marketing Strategy Recommendation

[0051] The trained model is used to process new samples to achieve marketing strategy recommendations. In one embodiment, the collected raw data is input into a trained Extreme Learning Machine algorithm based on fractional derivatives for classification, thereby obtaining classification results. In this embodiment, the classified categories include: promotional strategies, new product promotion strategies, high-end product strategies, etc. (the same as the labeled categories).

[0052] It should be noted that the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0053] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0054] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for training a marketing strategy recommendation model based on big data, characterized in that, Includes the following steps: Collect data from the product information management system, product manuals, and product detail pages on e-commerce platforms; A generative adversarial network algorithm based on sparse learning is used to generate samples, which expands the collected data and enhances the overall dataset. The expanded data is input into a pre-trained classifier model to obtain the classification result; The classifier model uses the extreme learning machine algorithm based on fractional derivatives as the classification algorithm; Based on the classification results, determine the marketing strategy outcomes.

2. The method for training a marketing strategy recommendation model based on big data according to claim 1, characterized in that: The training method for generative adversarial networks based on sparse learning is as follows: Initialize the generator and discriminator of the generative adversarial network. The weights of the generator and the discriminator are initialized with random numbers drawn from a Gaussian distribution. During each iteration of training, the generator extracts non-uniform feature information from the input noisy data and assigns dynamic weights to the feature information. The non-uniform features are weighted through a sparse learning-based mechanism. The generator generates data based on the input noise data and the weighted feature information; After receiving samples generated by the generator and samples from real data, the discriminator distinguishes them through a deep network layer and finally outputs the judgment result. The diversity of the generated samples is optimized by employing a diversity constraint term, which is calculated based on the information entropy of the feature distribution of the generated samples. During each round of training, the generator and discriminator are subjected to adversarial training. In each iteration of training, the generator optimizes the weights based on the sparsity of the features, gradually reducing the interference of redundant features; Update the weights of the generator and discriminator.

3. The method for training a marketing strategy recommendation model based on big data according to claim 1, characterized in that: The training method for the Limit Learning Machine algorithm based on fractional derivatives is as follows: The training data is input into the Extreme Learning Machine classifier, and the training data is mapped to a high-dimensional feature space through a nonlinear transformation. The weight matrix and bias terms of the extreme learning machine are initialized using the covariance matrix; In the hidden layers of the Extreme Learning Machine, the data is processed using a fractional-order differential strategy; Based on a fractional-order differential strategy, a self-similar curve-driven parameter learning method is used to update the weights of the extreme learning machine: The classification layer of the Extreme Learning Machine is the last layer, and the model output is converted into the predicted probability of each category through the Softmax function.

4. The method for training a marketing strategy recommendation model based on big data according to claim 1, characterized in that: This also includes preprocessing the collected data, including: Data cleaning removes meaningless symbols, stop words, and duplicate information, while also performing word segmentation on the text. The cleaned text data is vectorized using the term frequency-inverse term frequency method. The vectors of product attributes, descriptions, and functions are concatenated to form a comprehensive feature vector; The concatenated feature vectors are normalized, and the normalized feature vectors are used as input data for the machine learning model.

5. The method for training a marketing strategy recommendation model based on big data according to claim 1, characterized in that: Each data record contains the product's attributes, product description, and product features.

6. The method for training a marketing strategy recommendation model based on big data according to claim 2, characterized in that: Based on the sparse learning mechanism, the generator weights the non-uniform features, and the weighted features are: ； In the formula, These are dynamic learning parameters that represent the sparse weights of the features; It is a non-uniform characteristic; It is a sparsification function that forces some unimportant features to be compressed to zero; This represents the non-uniform characteristics after weighting.

7. The method for training a marketing strategy recommendation model based on big data according to claim 3, characterized in that: Fractional derivatives not only rely on traditional gradient information but also consider information from historical states. The calculation method is expressed as follows: ； In the formula, Denotes the fractional derivative. It is the order of the fractional differential; Let be the weight matrix of the extreme learning machine; This is a fractional weighting factor that adjusts the sensitivity of the fractional derivative to the current gradient. Let the loss function of the extreme learning machine be the gradient with respect to the weights. yes abbreviation, Indicates the first The loss function of the extreme learning machine in the next iteration with respect to the weight gradient, Let be the loss function of the extreme learning machine; Fractional dynamics representing time; This represents the integral operation, used to indicate the impact of past gradients on the current update; It is the step size of the historical iterations. It is different from The step size of the historical iterations; express differential.

Images