A Model-Based Method for Predicting Member Churn

Abstract

This invention relates to the field of user churn prediction technology and discloses a member churn prediction method based on model evaluation, comprising the following steps: extracting member behavior data, time series data, member feature data, and churn tag data from a member management system; performing data cleaning and normalization; and extracting time features; quantitatively evaluating the importance of features in member behavior data based on information entropy and conditional entropy calculations, and screening key features with conditional entropy values ​​below a preset threshold; constructing a synergistic effect function between features, and determining a dynamic weight allocation scheme for features through a game theory model; and based on the screened features and their corresponding weights. By quantifying the explanatory power of each feature for churn tags and screening key features, this method achieves the technical effect of improving model training efficiency and reducing interference from irrelevant features, solving the problem of model performance degradation due to insufficient feature importance evaluation.

Description

Technical Field

[0001] This invention relates to the field of user churn prediction technology, specifically a member churn prediction method based on model evaluation. Background Technology

[0002] With the rapid development of the internet economy, various online service platforms are proliferating, leading to increasingly fierce competition. How to effectively manage member resources and reduce member churn has become a core concern for many businesses. While existing member management systems can track member activity to some extent, they lack effective prediction of future member behavior, making it difficult for businesses to implement targeted strategies to retain members in advance.

[0003] Existing technologies for predicting member churn mainly involve extracting member behavior data and features, and then combining them with traditional machine learning models or deep learning models for modeling and prediction.

[0004] However, existing technologies for predicting member churn employ simplistic feature processing, failing to fully utilize time-series data and member characteristics. This results in models struggling to comprehensively capture member behavioral patterns. Feature optimization relies on simple correlation analysis, lacking dynamic weight adjustment capabilities and failing to adapt to changes in data distribution. Traditional machine learning models cannot capture time dependencies, and even with the introduction of time-series models, feature weight allocation and optimization mechanisms remain inadequate. Furthermore, model evaluation methods are overly static, neglecting comprehensive indicators, and optimization techniques are limited to parameter adjustments, lacking support for dynamic feature adjustment and real-time optimization, easily leading to high rates of missed and false positives. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a model-based method for predicting member churn. This method solves the problems of existing technologies, such as the single feature processing in member churn prediction, the failure to fully utilize time series data and member features, which makes it difficult for the model to comprehensively capture member behavior patterns; and the reliance on simple correlation analysis for feature optimization, which lacks the ability to dynamically adjust weights and is difficult to adapt to changes in data distribution.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a member churn prediction method based on model evaluation, comprising the following steps:

[0007] Member behavior data, time series data, member characteristic data, and churn tag data are extracted from the membership management system, and the data is cleaned, normalized, and time features are extracted.

[0008] Based on information entropy and conditional entropy calculations, the importance of features in member behavior data is quantitatively evaluated, and key features with conditional entropy values ​​below a preset threshold are selected.

[0009] Construct a synergistic effect function among features, and determine the dynamic weight allocation scheme of features through a game theory model;

[0010] Based on the selected features and their corresponding weights, a member churn prediction model is trained using a time series model to generate prediction output.

[0011] The model performance metrics are evaluated by comparing the predicted model output with the actual churned labels, and the feature weights and model parameters are dynamically adjusted based on the entropy increment.

[0012] Preferably, the steps of extracting member behavior data, time series data, member characteristic data, and churn tag data from the member management system, performing data cleaning and normalization processing, and extracting time features include:

[0013] Extract time interval features and periodic behavioral features from time series data, including calculating the time interval change rate of member behavior and extracting weekly and monthly periodic features of behavioral data;

[0014] Numerical features are normalized by using minimum and maximum normalization methods to map feature values ​​to the [0,1] interval in order to unify the data scale.

[0015] Preferably, the calculation based on information entropy and conditional entropy is used to calculate the information entropy of each feature, and to measure the complexity of the feature and the uniformity of the data distribution;

[0016] The explanatory power of each feature for predicting member churn is calculated based on conditional entropy. The smaller the conditional entropy value, the greater the contribution of the feature to the churn label.

[0017] Features whose conditional entropy value is lower than a preset threshold are selected as key features;

[0018] For time series data, features with significant dynamic trends are extracted by calculating the entropy increment of time features.

[0019] Preferably, the information entropy is defined as:

[0020]

[0021] The conditional entropy is defined as follows:

[0022]

[0023] Where H(Y|X) is the conditional entropy; P(x) i ,y j ) represents the value of feature X. i The value of the target variable Y j The joint probability; P(y) j |x i ) is the value of feature X as xi Under the given conditions, the value of the target variable Y is y. j The conditional probability; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y.

[0024] Preferably, the synergistic effect function between the constructed features is used to quantify the correlation between different features;

[0025] A game theory model is adopted with feature weight allocation as the optimization objective, and the dynamic weight allocation of features is solved through an iterative optimization algorithm.

[0026] Dynamically adjust feature weights to ensure that key features receive higher weights, thereby improving the model's ability to learn important features.

[0027] Preferably, the formula for the synergistic effect function is:

[0028]

[0029] Among them, G(X) i ,X j ) is the feature X i and X j The value of the synergistic effect function between them; H(Y|X) i ) is the feature X i The conditional entropy of the target variable Y; H(Y|X) j ) is the feature X j Conditional entropy of the target variable Y; smoothing factor ∈ (positive); X i and X j Y represents a feature from member behavior data, time series data, product data, or member characteristic data; Y is the target churn label.

[0030] Weight allocation satisfies the following constraints:

[0031]

[0032] Where w is the feature X i The weights are denoted by n; n is the number of features after filtering.

[0033] Preferably, the time series model is a long short-term memory network, and its network structure includes:

[0034] Input layer: Receives the feature set after feature optimization and its corresponding weights;

[0035] Hidden layer: Composed of long short-term memory network units, used to capture the time dependency of member behavior data;

[0036] Output layer: Outputs the predicted member churn probability value through a fully connected layer.

[0037] Preferably, the input features of the time series model are calculated using a weighted average:

[0038] The input feature vector is represented as:

[0039]

[0040] in, For feature X i The weight; is the i-th feature value at time point t in the time series; X t,1 X is the first feature value at time point t in the time series; n is the total number of input features. t This is the weighted input feature vector.

[0041] Preferably, the performance metrics of the evaluation model include the following steps:

[0042] Calculate the accuracy, recall, F1 score, and area under the curve of the prediction results;

[0043] Based on the entropy increment extracted from the time series, the feature weights are recalculated and the model parameters are adjusted.

[0044] The model should be updated regularly to ensure its robustness under different data distribution conditions.

[0045] Preferably, the present invention also provides a member churn prediction system based on model evaluation, comprising the following modules:

[0046] Data processing module: used for cleaning, normalizing, and extracting time-series features from member behavior data;

[0047] Feature optimization module: used to calculate information entropy and conditional entropy, and to filter key features;

[0048] Weight allocation module: Dynamically optimizes feature weights based on game theory;

[0049] Time series prediction module: Based on a long short-term memory network model, it trains and predicts the probability of member churn;

[0050] Model evaluation module: Evaluates prediction results and dynamically optimizes model parameters.

[0051] This invention provides a model-based method for predicting member churn. It offers the following advantages:

[0052] 1. This invention employs a feature optimization method combining information entropy and conditional entropy. By quantifying the explanatory power of each feature for lost labels, it filters out key features, achieving the technical effect of improving model training efficiency and reducing interference from irrelevant features. Compared to existing technologies that rely solely on human experience or simple statistical methods for feature selection, this invention solves the problem of model performance degradation caused by insufficient assessment of feature importance.

[0053] 2. This invention employs a game theory-based dynamic feature weight allocation mechanism. By constructing a feature synergy effect function and optimizing weight allocation, it maximizes the contribution of key features to model learning, thereby optimizing the model's learning ability. Compared to existing technologies with fixed or non-dynamically adjusted feature weights, this invention addresses the deficiency of model accuracy caused by changes in feature importance over time and data distribution.

[0054] 3. This invention uses Long Short-Term Memory (LSTM) networks to model time-series data and combines this with weighted optimized features to capture the time dependencies in member behavior data, achieving the technical effect of accurately predicting member churn trends. Compared to existing technologies where traditional machine learning models struggle to handle time dependencies, this invention solves the problems of low accuracy in member churn prediction and insufficient learning ability for time-series features.

[0055] 4. This invention employs a dynamic evaluation and optimization mechanism, evaluating the model's predictive performance through performance indicators and dynamically adjusting feature weights and model parameters by combining time series entropy increment, thereby enhancing the model's adaptability and robustness. Compared to existing static model training methods, this invention addresses the shortcomings of models being insensitive to changes in data distribution and exhibiting unstable performance in practical applications. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the present invention;

[0057] Figure 2 This is a system architecture diagram of the present invention; Detailed Implementation

[0058] The technical solution of the present invention will now be clearly and completely described 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.

[0059] Please see the appendix Figure 1This invention provides a model-based method for predicting member churn. Through a combination of multiple steps, including extracting and cleaning data from a member management system, optimizing and filtering features based on information entropy, constructing a feature synergy function and dynamically allocating weights using game theory, training a churn prediction model using a time series model, and dynamically evaluating optimization strategies, it achieves accurate prediction of member churn trends.

[0060] It should be noted that the high-quality data processing in step S1 provides input support for feature selection in S2. The selection results of S2 are further optimized for feature weights in S3. The optimized weights in S3 are directly used for time series model training in S4, while S5 has a feedback effect on the aforementioned steps through the evaluation and optimization mechanism.

[0061] like Figure 1 As shown, the member churn prediction method based on model evaluation of the present invention may include the following steps:

[0062] S1. Data extraction and preprocessing;

[0063] S2. Feature optimization and screening;

[0064] S3. Feature weight allocation;

[0065] S4. Training of the time series model;

[0066] S5. Model Evaluation and Optimization.

[0067] For step S1, this embodiment is divided into the following four steps:

[0068] Data extraction

[0069] Member behavior data includes page visit frequency, product clicks, number of times items were added to the shopping cart, number of completed transactions, and average transaction amount. For example, a member's browsing history over the past three months shows that they clicked on a certain type of product page 10 times and completed 3 purchases. This behavioral data is automatically recorded through a user behavior log system.

[0070] Time series data includes the date and time of specific operations, the time interval between events, and statistics on weekly or monthly active user behavior. Alternatively, time series data can also be used to extract daily transaction frequency or behavioral patterns to capture dynamic changes.

[0071] Member characteristic data includes static characteristics such as member registration time, level, and region, and dynamic characteristics such as loyalty score and last visit time.

[0072] Churn tag data: generated by rules based on no active behavior or historical annotations in the last three months, defined as binary tags: Y=1 indicates churn, Y=0 indicates no churn; in some embodiments, more complex tag rules can be used, such as combining tags with key business metrics.

[0073] Data cleaning

[0074] Missing value handling: For numerical features such as spending amount and access frequency, the mean imputation method is used; for categorical features such as member level and product category, the mode imputation method is used; for time series data, linear interpolation is used for imputation. For example, if a member's access frequency record for a certain month is missing, it can be replaced with the average of their access frequencies over the most recent three months.

[0075] Redundant feature removal: Remove features that are irrelevant or have low relevance to member churn prediction, such as fixed member identifiers that are completely static and have no temporal significance.

[0076] Normalization

[0077] To ensure dimensional consistency across different features, all numerical data are normalized using the following formula:

[0078]

[0079] Understandably, normalization effectively avoids model instability caused by excessively large or small feature value ranges.

[0080] Where x is the original feature value, min(x) is the minimum value of feature x in the entire dataset, max(x) is the maximum value of feature x in the entire dataset, and x′ is the normalized feature value.

[0081] Temporal Feature Extraction

[0082] Behavior time interval: Calculate the time interval between consecutive behaviors using timestamps.

[0083] Δt=t i+1 -t i

[0084] For example, if two consecutive purchase records are dated January 1, 2024 and January 10, 2024, the time interval is 9 days.

[0085] Time window statistics: This involves analyzing behavioral characteristics within a fixed time window (e.g., the last 30 days). For example, calculating a member's average spending and transaction frequency over the past month.

[0086] Where: t iGiven the timestamp of the i-th action, extract periodic behavioral features such as total monthly spending and weekly visit count; analyze the user's active behavioral phases (initial registration active period and stable period).

[0087] For step S2, this embodiment quantifies and evaluates the importance of each feature using information entropy and conditional entropy, eliminates redundant features, and retains key features that significantly contribute to churn prediction. Information entropy is used to measure the distributional complexity of a single feature. It is defined as the total amount of information uncertainty in the probability distribution of the feature value.

[0088] For example: if a feature has a uniform value range across all samples, then the entropy of that feature is high; conversely, if a feature is highly concentrated in certain specific values, then the entropy is low.

[0089] Conditional entropy is used to evaluate the explanatory power of features in predicting whether a target label will be churned; that is, the uncertainty of the target label given the feature's known characteristics. Features with lower conditional entropy indicate a stronger explanatory power for the target label. By combining information entropy and conditional entropy, and setting a conditional entropy threshold, features that meet the threshold requirements are selected. Features with conditional entropy below the set threshold are considered to contribute more strongly to member churn prediction. Furthermore, to dynamically capture member behavior trends, the entropy increment of time intervals is calculated for time-series features, extracting dynamic features with significant behavioral changes.

[0090] Information entropy measures the complexity of the distribution of feature data, and its calculation formula is:

[0091]

[0092] In some embodiments, the calculation of information entropy is applicable to discrete features, such as membership level and product category. It should be noted that for consumption amount and access frequency, information entropy can be calculated after discretizing them using a bucketing method.

[0093] Conditional entropy is used to quantify the explanatory power of features for lost labels:

[0094]

[0095] In one possible implementation, features with smaller conditional entropy values ​​have a stronger explanatory power for lost labels and are therefore more valuable for screening.

[0096] Feature selection rule: Set conditional entropy threshold δ H Filter features that meet the following conditions:

[0097]

[0098] As an option, the threshold δ HThe settings can be adjusted according to business needs. For example, when it is necessary to reduce the number of input features to the model, a stricter threshold can be selected to filter out the features with the greatest influence;

[0099] It should be noted that feature selection is not limited to conditional entropy; it can also incorporate the results of information entropy calculations. For example, features with high information entropy but low conditional entropy can be prioritized for retention, thereby ensuring that the input features possess both rich distribution information and high relevance to the lost labels.

[0100] Time series feature optimization:

[0101] Calculate the entropy increment of time series data:

[0102] ΔH t =H t -H t-1

[0103] To extract features that show significant dynamic changes, one possible approach is to prioritize retaining time features with large absolute values ​​of entropy increments, as these features often represent significant changes in the data.

[0104] Where H(Y|X) is the conditional entropy; P(x) i ,y j ) represents the value of feature X. i The value of the target variable Y j The joint probability; P(y) j |x i ) is the value of feature X as x i Under the condition that the value of the target variable Y is y j The conditional probability; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y; The set of features with significant dynamic changes selected; ΔH t δ represents the entropy increase of a time series feature at time t. H The entropy increment threshold for dynamic feature filtering.

[0105] For step S3, this embodiment employs a game theory-based weight optimization method to dynamically allocate weights to the selected feature set, enabling the model to better focus on key features. A feature synergy function measures the correlation between features and is used to optimize the importance ranking of features. By comparing the differences in the explanatory power of features for the target label, the complementary effects between different features are evaluated. The overall effect of feature weight allocation is optimized to ensure that the sum of weights is 1 and that the weights are non-negative. The purpose of dynamically optimizing the weights is to highlight key features while reducing the influence of redundant features. An iterative optimization algorithm (gradient descent) is used to calculate the feature weights, gradually converging towards the optimal weight allocation. The dynamically adjusted weights are directly used as input weighting factors for subsequent time series models.

[0106] Construction of the synergistic effect function, characteristic X i and X j The synergistic effect function is defined as:

[0107]

[0108] It should be noted that the above synergy effect function emphasizes the complementarity between features. The greater the difference, the smaller the value of the synergy effect. The model tends to allocate more weight to features with lower conditional entropy.

[0109] The optimization objective of weight allocation is to maximize the total contribution of the synergistic effect.

[0110]

[0111] Among them, w i For feature X i In one possible implementation, the weight allocation must satisfy the following constraints:

[0112]

[0113] Iterative optimization, using gradient descent to update weights:

[0114]

[0115] Among them, X i and X j G(X) represents a feature from member behavior data, time series data, product data, or member characteristic data; Y is the target churn tag; G(X) i ,X j ) is the feature X i and X j The value of the synergistic effect function between them; H(Y|X) i ) is the feature X i The conditional entropy of the target variable Y; H(Y|X) j) is the feature X j The conditional entropy of the target variable Y; ∈ is a positive smoothing factor; w is the feature weight vector; w i For feature X i The weights are η, where η is the learning rate.

[0116] In step S4, the filtered feature set, after weight optimization, forms the weighted input of the time series model. Weighting the input features helps the model more accurately capture the contribution of key features to churn prediction. A Long Short-Term Memory (LSTM) network is used as the core structure of the time series model. The time series model captures the dynamic characteristics of member behavior over time. Combining the filtered key features and their weights, the prediction model is trained to generate the probability output of member churn.

[0117] Input layer: Receives weighted feature inputs.

[0118] Hidden layer: Captures long-short-term dependencies in time series using long short-term memory network units.

[0119] Output layer: The hidden layer state is mapped to the member churn probability through a fully connected layer.

[0120] The loss function combines prediction error and feature weight optimization objectives, while also considering the deviation between the model's predictions and the true labels, as well as the regularization constraints of the feature weights on the model. This design ensures prioritized learning of key features while controlling model complexity.

[0121] The loss function combines cross-entropy and entropy regularization:

[0122]

[0123] Where L is the total loss function value; N is the total number of training samples; y i Let i be the true label of the i-th sample; Let be the predicted churn probability of the model for the i-th sample; and The value represents the logarithm of the predicted probability; λ is the regularization parameter; w j For feature X j The weights; H(Y|X) j ) is the feature X j The conditional entropy; n is the total number of features after filtering;

[0124] In one exemplary implementation, after training, the model generates a churn probability prediction for the input at each time point. To evaluate model performance, the following metrics can be used: accuracy (measures the proportion of correct predictions), recall (measures the model's ability to detect churned members), F1 score (combining precision and recall as a comprehensive evaluation metric), and area under the curve (reflects the model's ability to distinguish between churned and non-churned members).

[0125] For step S5, this embodiment evaluates the performance of the prediction model and, in conjunction with comparative analysis of actual churn labels, dynamically optimizes feature weights and model parameters to improve the overall prediction performance of the model. It should be noted that this step is logically closely related to the preceding steps and is a crucial step in ensuring the robustness of the model across different data distributions and scenarios.

[0126] As one implementation method, performance metrics are calculated based on the churn probability predicted by the model and the actual churn labels.

[0127] If it is found that the weights of certain features are insufficient, the weights are reallocated by increasing the entropy value, and the optimized features are then input into the time series model.

[0128] Based on the current model performance, different hyperparameter tuning strategies can be selected, such as increasing the number of hidden units in the time series model to improve the ability to capture time dependencies.

[0129] After the final optimized model, the evaluation process is repeated to ensure that the adjusted model has better predictive performance under different data distributions.

[0130] Please see the appendix Figure 2 The present invention also provides a member churn prediction system based on model evaluation. This system combines data processing, feature optimization, weight allocation, time series modeling and dynamic optimization techniques to ensure the accuracy and efficiency of member churn prediction.

[0131] This embodiment of the member churn prediction system based on model evaluation mainly includes the following functional modules: data processing module, feature optimization module, weight allocation module, time series prediction module, and model evaluation and optimization module. These modules work together to form a closed-loop system, continuously improving the prediction effect.

[0132] Data processing module:

[0133] This module is responsible for extracting and processing raw data from the membership management system to provide high-quality input for subsequent modules. Its main functions include: data extraction (reading member behavior data, time series data, member characteristic data, and churn tag data from the database); data cleaning (removing invalid data, filling in missing values, and eliminating outliers); data standardization (normalizing numerical features and unifying data units); and time feature extraction (generating time interval features and periodic behavior features).

[0134] Feature optimization module:

[0135] Data features are quantitatively evaluated and filtered based on information entropy and conditional entropy, eliminating redundant features and extracting key features. The implementation process is as follows: Calculate the information entropy H(X) of each feature to measure the complexity of the data distribution; calculate the conditional entropy H(Y|X) of the feature to evaluate its explanatory power for lost labels; set a threshold based on the conditional entropy to filter features that satisfy H(Y|X). i )≤δ H The key features are: calculating the entropy increment of time series features and extracting features with significant dynamic changes.

[0136] Weight allocation module

[0137] Game theory optimization models are used to assign weights to the selected features and dynamically adjust the importance of each feature. Implementation process: Construct a synergistic effect function G(X) among features. i ,X j This method measures the correlation between different features; defines the objective function for weight allocation and solves for the feature weights using gradient descent; and dynamically updates the weights to ensure that key features contribute more to model training.

[0138] Time series forecasting module:

[0139] This system predicts member churn probability using a time series model based on a weighted feature set. Main functions include: receiving a weighted, optimized feature set as input; constructing a time series model, including an input layer, hidden time series units, and an output layer; training the time series model using the weighted input features; and outputting the member churn probability.

[0140] Model evaluation and optimization module:

[0141] The performance of the prediction model is evaluated using multiple assessment metrics, and model parameters and feature weights are dynamically adjusted. The process involves: evaluating model performance using metrics such as accuracy, recall, F1 score, and AUC; dynamically adjusting feature weights based on entropy increments to optimize model input; and periodically updating the parameters of the time series model to improve its robustness under different data distributions.

[0142] The system modules in this embodiment are designed based on the execution flow of the above method embodiment. Their principles and technical effects are similar to those of the method embodiment, and will not be repeated here.

[0143] 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 model-based method for predicting member churn, characterized in that, Includes the following steps: Member behavior data, time series data, member characteristic data, and churn tag data are extracted from the membership management system, and the data is cleaned, normalized, and time features are extracted. Based on information entropy and conditional entropy calculations, the importance of features in member behavior data is quantitatively evaluated, and key features with conditional entropy values ​​below a preset threshold are selected. Construct a synergistic effect function among features, and determine the dynamic weight allocation scheme of features through a game theory model; Based on the selected features and their corresponding weights, a member churn prediction model is trained using a time series model to generate prediction output. The model performance metrics are evaluated by comparing the predicted model output with the actual churn labels, and the feature weights and model parameters are dynamically adjusted based on the entropy increase. The member behavior data includes page visit frequency, product clicks, number of times products are added to the shopping cart, number of completed transactions, and average transaction amount. The calculation based on information entropy and conditional entropy is used to calculate the information entropy of each feature, and to measure the complexity of the feature and the uniformity of the data distribution. The explanatory power of each feature for predicting member churn is calculated based on conditional entropy. The smaller the conditional entropy value, the greater the contribution of the feature to the churn label. Features whose conditional entropy value is lower than a preset threshold are selected as key features; For time series data, features with significant dynamic trends are extracted by calculating the entropy increment of time features; The synergistic effect function between the constructed features is used to quantify the correlation between different features; A game theory model is adopted with feature weight allocation as the optimization objective, and the dynamic weight allocation of features is solved through an iterative optimization algorithm. Dynamically adjust feature weights to ensure that key features receive higher weights, thereby improving the model's ability to learn important features; The formula for the synergistic effect function is: ； Among them, G(X) i ,X j ) is the feature X i and X j The value of the synergistic effect function between them; H(Y|X) i ) is the feature X i Conditional entropy of the target variable Y; H(Y|X) j ) is the feature X j Conditional entropy with respect to the target variable Y; A smoothing factor that is positive; X i and X j Y represents a feature from member behavior data, time series data, product data, or member characteristic data; Y is the target churn label. Weight allocation satisfies the following constraints: ； Among them, w i For feature X i The weights are denoted by n; n is the number of features after filtering. The optimization objective of weight allocation is to maximize the total contribution of the synergistic effect. ； Among them, w j For feature X j The weight.

2. The member churn prediction method based on model evaluation according to claim 1, characterized in that, The steps of extracting member behavior data, time series data, member characteristic data, and churn tag data from the member management system, performing data cleaning and normalization, and extracting time features include: Extract time interval features and periodic behavioral features from time series data, including calculating the time interval change rate of member behavior and extracting weekly and monthly periodic features of behavioral data; Numerical features are normalized by using minimum and maximum normalization methods to map feature values ​​to the [0,1] interval in order to unify the data scale.

3. The member churn prediction method based on model evaluation according to claim 1, characterized in that, The definition of information entropy is: ； The conditional entropy is defined as follows: ； Where H(Y|X) is the conditional entropy; P(x) i ,y j The value of feature X is X. i The value of the target variable Y j The joint probability; P(y) j |x i ) is the value of feature X. i Under the given conditions, the value of the target variable Y is y. j The conditional probability; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y.

4. The member churn prediction method based on model evaluation according to claim 1, characterized in that, The time series model is a long short-term memory network, and its network structure includes: Input layer: Receives the feature set after feature optimization and its corresponding weights; Hidden layer: Composed of long short-term memory network units, used to capture the time dependency of member behavior data; Output layer: Outputs the predicted member churn probability value through a fully connected layer.

5. The member churn prediction method based on model evaluation according to claim 4, characterized in that, The input features of the time series model are calculated using a weighted average: The input feature vector is represented as: ； in, For feature X i The weight; is the i-th feature value at time point t in the time series; X t,1 X is the first feature value at time point t in the time series; n is the total number of input features. t This is the weighted input feature vector.

6. The member churn prediction method based on model evaluation according to claim 1, characterized in that, The performance metrics of the evaluation model include the following steps: Calculate the accuracy, recall, F1 score, and area under the curve of the prediction results; Based on the entropy increment extracted from the time series, the feature weights are recalculated and the model parameters are adjusted. The model should be updated regularly to ensure its robustness under different data distribution conditions.

7. A model-based churn prediction system, applied to the model-based churn prediction method described in claims 1-6, characterized in that, The model-based member churn prediction system includes the following modules: Data processing module: used for cleaning, normalizing, and extracting time-series features from member behavior data; Feature optimization module: used to calculate information entropy and conditional entropy, and to filter key features; Weight allocation module: Dynamically optimizes feature weights based on game theory; Time series prediction module: Based on a long short-term memory network model, it trains and predicts the probability of member churn; Model evaluation module: Evaluates prediction results and dynamically optimizes model parameters.

Images