Communication user loss prediction system and method based on improved generative adversarial network and heterogeneous model integration

Abstract

The invention discloses a communication user loss prediction method based on integration of an improved generative adversarial network and a heterogeneous model, and the method comprises the steps: data preprocessing and feature engineering: preprocessing user data, and constructing a feature set to provide an input basis for a subsequent generative adversarial network and a prediction model; designing an improved generative adversarial network, training a GAN model based on a minimum-maximum game, and laying a model foundation for generating a high-fidelity synthesis sample in a data enhancement stage; performing data enhancement and reconstruction, generating lost user samples in the original data set by using an improved GAN model, and fusing the generated samples with the original data set to form a new balanced data set for subsequent prediction model training; building a user loss prediction model, training an XGBoost sub-model and a neural network sub-model based on the obtained balanced data set, integrating the XGBoost sub-model and the neural network sub-model through weighted fusion, and building the user loss prediction model; and applying the integrated prediction model to carry out loss prediction analysis on the new user data.

Description

Technical Field

[0001] This invention belongs to the field of big data analysis and artificial intelligence in the communications industry, specifically involving a communication user churn prediction system and method based on the integration of improved generative adversarial networks and heterogeneous models. Background Technology

[0002] This section provides only background information relevant to this disclosure and is not necessarily prior art.

[0003] In the telecommunications industry, customer churn has always been a core business challenge for operators. Traditional methods for predicting customer churn are mainly based on statistical learning models such as logistic regression, support vector machines (SVM), random forests (RF), and XGBoost. These methods have achieved certain results in feature extraction and classification prediction, but their predictive performance drops significantly when faced with customer churn data with imbalanced category distributions.

[0004] In existing technologies, resampling techniques (such as SMOTE) are commonly used to address class imbalance by synthesizing minority class samples to balance the dataset. However, these methods are prone to introducing noise when dealing with high-dimensional heterogeneous data, causing the generated samples to deviate from the true data pattern and affecting the stability and accuracy of the classification model.

[0005] Furthermore, traditional models struggle to effectively capture the inherent patterns of complex user behavior characteristics, resulting in limited predictive performance. Therefore, there is an urgent need for a user churn prediction method that can address the class imbalance problem and improve prediction accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a communication user churn prediction system and method based on an improved generative adversarial network and heterogeneous model integration, thereby solving the performance degradation problem of existing user churn prediction models under conditions of class distribution imbalance. By introducing an improved generative adversarial network and combining it with heterogeneous model integration technology, the synergistic evolution of data reconstruction and prediction optimization is achieved.

[0007] A method for predicting communication user churn based on the integration of an improved generative adversarial network and a heterogeneous model includes the following steps:

[0008] Step 1, data preprocessing and feature engineering, includes:

[0009] Step 1-1, missing value handling: The initial dataset is checked for missing values ​​to find missing records, and the K-nearest neighbor algorithm is used for interpolation to fill in the missing records.

[0010] Step 1-2: Duplicate value processing. Perform data deduplication to ensure data uniqueness. If duplicate samples exist, delete redundant entries.

[0011] Steps 1-3 involve data type conversion: mapping and converting binary categorical variables, and performing one-hot encoding on multi-categorical variables.

[0012] Steps 1-4: Outlier detection and cleaning. For numerical features, the interquartile range (IQR) method is used for outlier identification. The first quartile (Q1) and the third quartile (Q3) are calculated. The IQR is defined as Q3-Q1, and the outlier boundary is defined as follows:

[0013]

[0014]

[0015] Remove abnormal samples that exceed the boundary.

[0016] Steps 1-5 involve redundant feature removal. First, irrelevant identifier fields are deleted. Then, low-variance features are filtered based on analysis of variance. Finally, highly collinear features are removed through Pearson correlation coefficient analysis.

[0017] Steps 1-6: Dimensional compression. Principal component analysis (PCA) is used to perform a linear transformation on the feature space, retaining 95% of the original data variance, and compressing the feature space to k dimensions (k is automatically calculated by PCA).

[0018] Steps 1-7: Data standardization using the Z-score standardization method, mathematically expressed as:

[0019]

[0020] in, For sample feature values, The characteristic mean, The characteristic standard deviation is denoted as .

[0021] Step 2, Improved Generative Adversarial Network Design, such as Figure 1 As shown in the diagram of the generative adversarial network architecture, its core consists of two deep neural networks, including:

[0022] generator : Receive random noise vector As input, synthetic data is generated through nonlinear transformation. The goal is to deceive the discriminator.

[0023] Discriminator : Receive real data Or generate data It outputs a probability estimate of the authenticity of the sample, with the goal of accurately distinguishing the data source.

[0024] Both follow the minimax game principle, and the objective function can be expressed as:

[0025] .

[0026] Discriminator Maximize the objective function, generator Minimize the objective function, where, Represents the true data distribution. Indicates the prior noise distribution. and These represent the probabilities that the discriminator outputs real data and the production data are real samples, respectively.

[0027] Step 2-1, Generator Employing a deep residual structure, the network architecture is as follows: Figure 2 As shown, the input layer maps the latent space vectors to the deep feature space through a linear transformation. A multi-head attention mechanism is used to establish global dependencies between features along the sequence dimension. Subsequently, a residual module is connected. Each layer includes a linear transformation, layer normalization (LayerNorm), and the LeakyReLU activation function. The residual connections use a skip connection approach. Finally, through nonlinear transformation of the hidden layer, the synthesized sample is output after Tanh activation;

[0028] Step 2-2, Discriminator A three-layer fully connected architecture is adopted. To avoid the gradient vanishing problem, the hidden layer uses the LeakyReLU activation function, and the output layer uses the Sigmoid function to implement probability mapping.

[0029] Steps 2-3: Model training employs a hybrid loss function of Wasserstein distance and Maximum Mean Discrepancy (MMD). The algorithm training uses an alternating optimization strategy, fixing the generator parameters and updating the discriminator parameters.

[0030]

[0031] With discriminator parameters fixed, update generator parameters:

[0032]

[0033] in , These are the weight coefficients for the gradient penalty term and the MMD regularization term, respectively. Dynamically adjust using cosine annealing strategy. This is the total training cycle.

[0034] .

[0035] Step 3, Data Augmentation and Reconstruction: Using the improved GAN model, new churn user data samples are generated from the original dataset. These new samples are then merged with the original dataset to form a new balanced dataset. This dataset contains the same number of churn and non-churn user samples, which are then used for training the subsequent prediction model.

[0036] Step 4, building a user churn prediction model, including:

[0037] Step 4-1, XGBoost sub-model training, involves iteratively generating a sequence of decision trees to progressively fit the target variable. The construction of each tree is based on the feature information gain maximization criterion. XGBoost iteratively optimizes the prediction results using an additive training strategy.

[0038]

[0039] Its objective function is:

[0040]

[0041] in The total number of samples in the training dataset, The total number of decision trees, and the regularization term. , The number of leaf nodes in the tree. The weight of the leaf node. and These are hyperparameters that penalize the number of leaf nodes and the leaf weight, respectively.

[0042] Step 4-2: Training the neural network sub-model. This network adopts a fully connected architecture design, gradually extracting deep patterns of user behavior through hierarchical feature transformation. The network inserts batch normalization layers (BatchNorm) and dropout layers between hidden layers, and uses class-weighted cross-entropy loss as the loss function.

[0043]

[0044] in, The total number of samples, For the first One sample, For the first The label corresponding to each sample Predict samples for the model The probability of belonging to the positive class. For class weights, the AdamW optimizer is used to update the parameters:

[0045]

[0046] in, For the first Step model parameters, The initial learning rate, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, This is the weight decay coefficient. For numerically stable terms, the weights The range is [0,1], determined by the number of positive and negative samples in the dataset, and is not updated during training.

[0047] Step 4-3: Heterogeneous Model Integration. A weighted fusion prediction model is constructed, integrating the XGBoost model and the neural network sub-model. A weighted fusion mechanism is established, with the optimization objective of maximizing the F1-score on the validation set. The optimal weight ratio between XGBoost and the deep neural network is determined through grid search, with a search step size of 0.1 and a search range of [0.3, 0.7]. These are the weights for the neural network and XGBoost, respectively. The formula is as follows:

[0048]

[0049] in It is the prediction result after fusion. These are the prediction results of the neural network sub-model and the XGBoost model, respectively. To avoid excessive computational overhead, the search range was determined empirically, and to avoid the weight of a single model being too large or too small, XGBoost is more cautious in distinguishing positive samples, resulting in higher accuracy and precision, but lower recall. The neural network has a slightly stronger ability to capture positive samples, resulting in higher recall, but lower accuracy and precision.

[0050] Step 5: Apply the integrated prediction model to perform churn prediction analysis on new user data.

[0051] A communication user churn prediction system based on the integration of improved generative adversarial networks and heterogeneous models is characterized by using a communication user churn prediction method based on the integration of improved generative adversarial networks and heterogeneous models.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] Addressing the class imbalance problem: High-quality churn user data samples are generated through an improved GAN model, significantly improving the class distribution imbalance problem and increasing the coverage of minority class samples.

[0054] High-quality generated samples: Compared with the traditional SMOTE method, the GAN model of this invention improves the recall rate while maintaining the discriminative features of the generated samples, and avoids the noise problem caused by oversampling.

[0055] The model is robust: By weighted fusion of neural networks and XGBoost, the advantages of deep learning and traditional statistical models are effectively integrated, improving the robustness and generalization ability of the model. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the generative adversarial network of the present invention.

[0057] Figure 2 This is a diagram of the generator architecture of the present invention.

[0058] Figure 3 This is a diagram of the discriminator architecture of the present invention.

[0059] Figure 4 This is a flowchart of the present invention. Detailed Implementation

[0060] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0061] Example 1:

[0062] like Figure 4 As shown in the figure, this embodiment provides a method for predicting communication user churn, which is performed according to the following steps:

[0063] Step 1: Collect communication user churn data to obtain a user churn dataset;

[0064] In this embodiment, the user churn data comes from publicly available data on the Kaggle platform. This dataset was collected from a real business scenario of a telecommunications company and contains 7,043 user records and 21 valid fields. As shown in Table 1, the data dimensions cover user attributes (gender, family situation, etc.), account information (usage duration, contract type, payment method, etc.), service characteristics (value-added service subscriptions, etc.), and the core prediction target—user churn status label (Churn Label), forming a complete user profile system. The dataset contains 5,174 users who have not churned and 1,869 users who have churned. Through the coupling relationship between feature variables and binary classification target variables, a data foundation is provided for the construction and validation of a user churn early warning model for the telecommunications industry.

[0065] Table 1. Field Descriptions for User Churn Dataset

[0066]

[0067] Step 2: Perform data preprocessing and feature engineering on the aforementioned communication user churn dataset to obtain a preprocessed dataset for training the user churn prediction model; the preprocessing includes missing value handling, duplicate value handling, data type conversion, outlier detection and cleaning, redundant feature removal, dimensionality compression, and data standardization.

[0068] Because the original churned user data set may contain missing values, duplicate samples, unstructured data, outliers, and redundant features, systematic data cleaning and feature optimization are necessary to improve the training efficiency and prediction accuracy of the model. This includes the following steps:

[0069] Step 2-1: Missing value handling. The initial dataset is checked for missing values. After missing records are found, the K-Nearest Neighbors (KNN) algorithm is used for interpolation to fill in the missing records.

[0070] In this embodiment, the missing values ​​are filled in by calculating the feature similarity between the missing sample and its K nearest neighbor samples, so as to preserve data integrity and reduce information loss.

[0071] Step 2-2: Duplicate value handling. Perform data deduplication to ensure data uniqueness; if duplicate samples exist, delete redundant entries.

[0072] In this embodiment, duplicate records are identified and deleted by comparing the customer ID field, ensuring the uniqueness of each user's data.

[0073] Steps 2-3: Data type conversion, mapping and conversion of binary categorical variables, and one-hot encoding of multi-categorical variables.

[0074] In this embodiment, binary categorical variables such as “Gender”, “Partner”, and “Dependents” are mapped to 0 / 1 form, and multi-categorical variables such as “InternetService”, “Contract”, and “PaymentMethod” are one-hot encoded to adapt to the model input format.

[0075] Steps 2-4: Outlier Detection and Cleaning. For numerical features, the Interquartile Range (IQR) method is used for outlier identification. The first quartile (Q1) and the third quartile (Q3) are calculated. The IQR is defined as Q3-Q1, and the outlier boundary is defined as follows:

[0076]

[0077]

[0078] In this embodiment, abnormal samples that exceed the boundary range are removed to improve data quality and model stability.

[0079] Steps 2-5: Redundant feature removal. First, delete irrelevant identifier fields, then filter low variance features based on analysis of variance, and finally remove highly collinear features through Pearson correlation coefficient analysis.

[0080] In this embodiment, firstly, identifier fields such as "customerID" that are irrelevant to the prediction target are deleted. Then, variance analysis is used to filter out features that have a significant impact on user churn status. Finally, Pearson correlation coefficient analysis is used to remove redundant features with a correlation higher than 0.9, so as to reduce model complexity and improve generalization ability.

[0081] Steps 2-6: Dimensional compression. Principal component analysis (PCA) is used to perform a linear transformation on the feature space, retaining 95% of the original data variance, and compressing the feature space to k dimensions (k is automatically calculated by PCA).

[0082] In this embodiment, the original feature space is compressed to k dimensions using the PCA dimensionality reduction method, where k is automatically determined when the cumulative variance contribution rate reaches 95%, thereby reducing computational overhead while retaining the main information.

[0083] Steps 2-7: Data standardization, using the Z-score standardization method, its mathematical expression is:

[0084]

[0085] in, For sample feature values, The characteristic mean, The characteristic standard deviation;

[0086] In this embodiment, Z-score standardization is used to make the mean 0 and the standard deviation 1, so as to eliminate the difference in dimensions and improve the convergence speed and stability of the model.

[0087] Step 3: Use the improved generative adversarial network model to perform data augmentation and reconstruction on the churned user samples in the original dataset. This specifically includes the following steps:

[0088] Step 3-1: Improved design and training of generative adversarial networks, such as... Figure 1 As shown in the diagram of the generative adversarial network architecture, its core consists of two deep neural networks, including:

[0089] generator : Receive random noise vector As input, synthetic data is generated through nonlinear transformation. The goal is to deceive the discriminator;

[0090] Discriminator : Receive real data Or generate data It outputs a probability estimate of the authenticity of the sample, with the goal of accurately distinguishing the data source.

[0091] Both follow the minimax game principle, and the objective function can be expressed as:

[0092]

[0093] Among them, generator Employing a deep residual structure, the network architecture is as follows: Figure 2 As shown, the input layer maps the latent space vectors to the deep feature space through a linear transformation. A multi-head attention mechanism is used to establish global dependencies between features along the sequence dimension. Subsequently, a residual module is connected. Each layer includes a linear transformation, layer normalization (LayerNorm), and the LeakyReLU activation function. The residual connections use a skip connection approach. Finally, through a nonlinear transformation of the hidden layer, the synthesized sample is output after Tanh activation. Discriminator like Figure 3 As shown, a three-layer fully connected architecture is adopted. To avoid the gradient vanishing problem, the hidden layer uses the LeakyReLU activation function, and the output layer uses the Sigmoid function to implement probability mapping.

[0094] In this embodiment, model training is implemented using the PyTorch deep learning framework, where modules such as multi-head self-attention mechanism, layer normalization, and LeakyReLU activation function are obtained by calling the corresponding functions of the framework.

[0095] Based on the above design, an improved generative adversarial network was obtained, and the generative adversarial network was trained using data.

[0096] Optionally, in step 3, the following strategy is adopted when training the generative adversarial network:

[0097] The model training uses a hybrid loss function of Wasserstein distance and maximum mean difference, and the algorithm training employs an alternating optimization strategy, fixing the generator parameters and updating the discriminator parameters.

[0098]

[0099] With discriminator parameters fixed, update generator parameters:

[0100]

[0101] in , These are the weight coefficients for the gradient penalty term and the MMD regularization term, respectively. Dynamically adjust using cosine annealing strategy. This is the total training cycle.

[0102]

[0103] Step 3-2: Data Augmentation and Reconstruction. Utilize the trained generator to generate high-quality synthetic churn user data samples. Merge the generated synthetic churn user samples with the samples in the original dataset to form a new balanced dataset. This balanced dataset contains both the original churn user samples and the generated churn user samples, which are used to train the subsequent prediction model, thereby improving the model's performance and generalization ability in the churn user prediction task.

[0104] Step 4: Use the user churn prediction model described above to predict churned users. This specifically includes the following steps:

[0105] Step 4-1: XGBoost sub-model training. This involves iteratively generating a sequence of decision trees to progressively fit the target variable. The construction of each tree is based on the feature information gain maximization criterion. XGBoost uses an additive training strategy to iteratively optimize the prediction results.

[0106]

[0107] Its objective function is:

[0108]

[0109] in The total number of samples in the training dataset, The total number of decision trees, and the regularization term. , The number of leaf nodes in the tree. The weight of the leaf node. and These are hyperparameters that penalize the number of leaf nodes and the leaf weight, respectively.

[0110] Step 4-2: Training the neural network sub-model. This network adopts a fully connected architecture design, gradually extracting deep patterns of user behavior through hierarchical feature transformation. The network inserts batch normalization layers (BatchNorm) and random deactivation layers (Dropout) between hidden layers, and uses class-weighted cross-entropy loss as the loss function.

[0111]

[0112] in, The total number of samples, For the first One sample, For the first The label corresponding to each sample Predict samples for the model The probability of belonging to the positive class. For class weights, the AdamW optimizer is used to update the parameters:

[0113]

[0114] in, For the first Step model parameters, The initial learning rate, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, This is the weight decay coefficient. It is a numerically stable term.

[0115] Step 4-3: Heterogeneous Model Integration. A weighted fusion prediction model is constructed, integrating the XGBoost model and the neural network sub-model. A weighted fusion mechanism is established, with the optimization objective of maximizing the F1-score on the validation set. The optimal weight ratio between XGBoost and the deep neural network is determined through grid search, with a search step size of 0.1 and a search range of [0.3, 0.7]. and The weights for the neural network and XGBoost are shown in the following formula:

[0116]

[0117] in It is the prediction result after fusion. These are the prediction results from the neural network sub-model and the XGBoost model, respectively.

[0118] In this embodiment, the user churn prediction method provided is compared with nine classic algorithms in the prior art, including Random Forest, Gradient Boosting, AdaBoost, XGBoost, Support Vector Machine (SVM), Logistic Regression, Naive Bayes, K-Nearest Neighbors (KNN), and Decision Tree. The experiment employs a four-dimensional evaluation system, including accuracy, precision, recall, and F1 score, to comprehensively measure the model's predictive performance.

[0119] Table 2 Performance comparison of different algorithms on the test set

[0120]

[0121] As shown in Table 2, the metrics demonstrate that the algorithm proposed in this paper is in a leading position, proving that the model can more effectively identify potential churned users, which has important practical value for enterprises to implement customer retention strategies.

[0122] This invention constructs an improved generative adversarial network model: including a collaborative structure of multi-head self-attention mechanism and residual connections, to generate high-quality churned user data samples. This increases the number of minority class samples and alleviates the problem of class distribution imbalance.

[0123] Establish a user churn prediction model: Construct a weighted fusion model of neural networks and XGBoost, combining the advantages of deep learning and statistical models to improve prediction performance.

[0124] Real-time churn prediction system: including a complete process of data preprocessing, GAN data augmentation, and model integration prediction, used for churn prediction in the telecommunications industry.

[0125] This invention provides an idea and method for a communication user churn prediction system and method based on the integration of an improved generative adversarial network and a heterogeneous model. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting communication user churn based on the integration of an improved generative adversarial network and a heterogeneous model, characterized in that, The specific steps include the following: Step 1, Data Preprocessing and Feature Engineering: Preprocess user data to construct a feature set, providing the input basis for subsequent generative adversarial networks and prediction models; Step 2: Design an improved generative adversarial network and train a GAN model based on minimax game to lay the model foundation for generating high-fidelity synthetic samples in the data augmentation stage. Step 3, Data Augmentation and Reconstruction: Use the improved GAN model obtained in Step 2 to generate churned user samples from the original dataset, and merge the generated samples with the original dataset to form a new balanced dataset for subsequent prediction model training. Step 4: Building the user churn prediction model. Based on the balanced dataset obtained in Step 3, train the XGBoost sub-model and the neural network sub-model, and integrate the two through weighted fusion to build the user churn prediction model. Step 5: Apply the integrated prediction model to perform churn prediction analysis on new user data.

2. The method according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1, missing value handling: The initial dataset is checked for missing values ​​to identify missing records, and the K-nearest neighbor algorithm is used for interpolation to fill in the missing records. Step 1-2: Duplicate value processing. Perform data deduplication to ensure data uniqueness. If duplicate samples exist, delete redundant entries. Steps 1-3: Data type conversion, mapping conversion for binary categorical variables, and one-hot encoding for multi-categorical variables; Steps 1-4: Outlier detection and cleaning. For numerical features, the interquartile range (IQR) method is used for outlier identification. The first quartile (Q1) and the third quartile (Q3) are calculated. The IQR is defined as Q3-Q1, and the outlier boundary is defined as follows: Remove abnormal samples that exceed the boundary; Steps 1-5: Redundant feature removal, deletion of irrelevant identifier fields, filtering of low variance features based on analysis of variance, and removal of highly collinear features through Pearson correlation coefficient analysis; Steps 1-6: Dimensional compression. Principal component analysis (PCA) is used to perform a linear transformation on the feature space, retaining 95% of the original data variance, and compressing the feature space to k dimensions, where k is calculated by PCA. Steps 1-7: Data standardization using the Z-score standardization method, mathematically expressed as: in, For sample feature values, The characteristic mean, denoted as the characteristic standard deviation, and i as the label.

3. The method according to claim 2, characterized in that, The generative adversarial network in step 2 consists of two deep neural networks, including: Generator G: Receives a random noise vector z as input and generates synthetic data through nonlinear transformation. The goal is to deceive the discriminator; Discriminator D: Receives real data Or generate data The output sample authenticity estimate aims to accurately distinguish the data source. Both follow the minimax game principle, and the objective function is expressed as: Discriminator Maximize the objective function, generator Minimize the objective function, where, Represents the true data distribution. Indicates the prior noise distribution. and These represent the probabilities that the discriminator outputs real data and the production data are real samples, respectively.

4. The method according to claim 3, characterized in that, Step 2 specifically includes the following steps: Step 2-1, Generator A deep residual structure is adopted. The input layer maps the latent space vector to the deep feature space through linear transformation. A multi-head self-attention mechanism is used to establish global dependencies between features in the sequence dimension. The residual module is connected. Each layer contains linear transformation, layer normalization and LeakyReLU activation function. The residual connection adopts a skip connection method. Through the nonlinear transformation of the hidden layer, the synthetic sample is output after Tanh activation. Step 2-2, Discriminator A three-layer fully connected architecture is adopted, with the hidden layer using the LeakyReLU activation function and the output layer using the Sigmoid function to implement probability mapping; Steps 2-3: Model training uses a hybrid loss function of Wasserstein distance and maximum mean difference. Algorithm training employs an alternating optimization strategy, fixing the generator parameters and updating the discriminator parameters. With discriminator parameters fixed, update generator parameters: in , These are the weight coefficients for the gradient penalty term and the MMD regularization term, respectively. Dynamically adjust using cosine annealing strategy. For the total training cycle; 。 5. The method according to claim 4, characterized in that, Step 3 specifically involves generating new churned user data samples from the original dataset using the improved GAN model. These samples are then merged with the original dataset to form a new balanced dataset, in which the number of churned and non-churned user samples is equal, which is then used for subsequent prediction model training.

6. The method according to claim 5, characterized in that, Step 4 specifically includes the following steps: Step 4-1, XGBoost sub-model training: A sequence of decision trees is generated iteratively to gradually fit the target variable. The construction of each tree is based on the feature information gain maximization criterion. XGBoost iteratively optimizes the prediction results using an additive training strategy. Its objective function is: in The total number of samples in the training dataset, The total number of decision trees, and the regularization term. , The number of leaf nodes in the tree. The weight of the leaf node. and These are hyperparameters that penalize the number of leaf nodes and the leaf weight, respectively. Step 4-2: Training the neural network sub-model. The neural network adopts a fully connected architecture design, gradually extracting deep patterns of user behavior through hierarchical feature transformation. Batch normalization layers and random deactivation layers are inserted between hidden layers. Class-weighted cross-entropy loss is used as the loss function. in, The total number of samples, For the first One sample, For the first The label corresponding to each sample Predict samples for the model The probability of belonging to the positive class. For class weights, the AdamW optimizer is used to update the parameters: in, For the first Step model parameters, The initial learning rate, This is a first-moment estimate of the gradient. For the second moment estimation of the gradient, This is the weight decay coefficient. It is a numerically stable term; Step 4-3: Heterogeneous model integration. Construct a weighted fusion prediction model of the XGBoost model and the neural network sub-model, and establish a weighted fusion mechanism.

7. The method according to claim 6, characterized in that, The fusion method in step 4-3 is as follows: With maximizing the F1-score on the validation set as the optimization objective, the optimal weight ratio between XGBoost and the deep neural network is determined through grid search, with a search step size of 0.1 and a search range of [0.3, 0.7]. and The weights for the neural network and XGBoost are shown in the following formula: in It is the prediction result after fusion. , These are the prediction results from the neural network sub-model and the XGBoost model, respectively. .

8. A communication user churn prediction system based on the integration of an improved generative adversarial network and a heterogeneous model, characterized in that, Use any one of the communication user churn prediction methods based on the integration of improved generative adversarial networks and heterogeneous models as described in claims 1 to 7.

Images