Member loss prediction method based on model evaluation

Abstract

The invention relates to the technical field of user loss prediction, and discloses a member loss prediction method based on model evaluation, comprising the following steps: extracting member behavior data, time sequence data, member feature data and loss label data from a member management system, performing data cleaning and normalization processing, and extracting time features; performing quantitative evaluation on feature importance in the member behavior data based on information entropy and conditional entropy calculation, and screening key features with conditional entropy lower than a preset threshold value; constructing a synergistic effect function between the features, and determining a dynamic weight distribution scheme of the features through a game theory model; and based on the screened features and the corresponding weights. The key features are screened out by quantifying the explanation ability of each feature for the lost labels, the technical effects of improving the model training efficiency and reducing irrelevant feature interference are achieved, and the problem that the model performance is reduced due to insufficient feature importance evaluation is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of user churn prediction, and in particular to a member churn prediction method based on model evaluation. Background Art

[0002] With the rapid development of the Internet economy, various online service platforms are increasing, and competition is becoming increasingly fierce. How to effectively manage member resources and reduce member churn has become one of the core issues that many companies are concerned about. Although the existing member management system can count member activities to a certain extent, it lacks effective prediction of members' future behavior, making it difficult for companies to adopt targeted strategies in advance to retain members; Existing technologies mainly extract member behavior data and features and combine traditional machine learning models or deep learning models to perform modeling and prediction in member churn prediction; However, the existing technology has a single feature processing when predicting member churn, and does not fully utilize time series data and member characteristics, which makes it difficult for the model to fully capture the rules of member behavior; feature optimization relies on simple correlation analysis, lacks the ability to dynamically adjust weights, and is difficult to adapt to changes in data distribution; traditional machine learning models cannot capture time dependencies, and even if time series models are introduced, feature weight allocation and optimization mechanisms are still imperfect. In addition, the model evaluation method is too static, ignores comprehensive indicators, and the optimization method is limited to parameter adjustment. There is a lack of support for dynamic adjustment and real-time optimization of features, which can easily lead to high rates of missed judgments and misjudgments. Summary of the invention

[0003] In view of the shortcomings of the existing technology, the present invention provides a member churn prediction method based on model evaluation, which solves the problems that the existing technology has single feature processing in member churn prediction, does not fully utilize time series data and member characteristics, resulting in the model being difficult to fully capture the rules of member behavior; feature optimization relies on simple correlation analysis, lacks the ability to dynamically adjust weights, and is difficult to adapt to changes in data distribution.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: a member churn prediction method based on model evaluation, comprising the following steps: Extract member behavior data, time series data, member feature data, and churn label data from the member management system, perform data cleaning and normalization, and extract time features; 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 the preset threshold are screened; Construct the synergy effect function between features and determine the dynamic weight distribution scheme of features through the game theory model; Based on the screened features and corresponding weights, the member churn prediction model is trained using a time series model to generate prediction outputs. By comparing the prediction model output with the actual churn labels, the model performance indicators are evaluated, and the feature weights and model parameters are dynamically adjusted according to the entropy value.

[0005] Preferably, the steps of extracting member behavior data, time series data, member feature data and churn label data from the member management system, performing data cleaning and normalization processing, and extracting time features include: Extracting time interval features and periodic behavior features from time series data, including calculating the time interval change rate of member behavior and extracting weekly and monthly periodic features of behavior data; The numerical features are normalized and the minimum and maximum normalization methods are used to map the feature values ​​to the [0,1] interval to unify the data scale.

[0006] Preferably, the information entropy and conditional entropy calculation is used to calculate the information entropy of each feature, measuring the complexity of the feature and the uniformity of data distribution; The explanatory power of each feature for member churn prediction is calculated based on conditional entropy. The feature with a smaller conditional entropy value contributes more to the churn label. The features whose entropy values ​​are lower than the preset threshold are selected as key features; For time series data, the features with significant dynamic change trends are extracted by calculating the entropy value of time features.

[0007] Preferably, the information entropy is defined as: The conditional entropy is defined as: Among them, H(Y|X) is the conditional entropy; P(x i ,y j ) is the value x of feature X i and the value y of the target variable Y j The joint probability of j ∣x i ) is the value of feature X when i Under the condition that j The conditional probability of ; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y.

[0008] Preferably, the synergistic effect function between the constructed features is used to quantify the correlation between different features; The game theory model is used to take feature weight allocation as the optimization goal, and the dynamic weight allocation of features is solved through an iterative optimization algorithm; the feature weights are dynamically adjusted to ensure that key features obtain higher weights, thereby improving the model's ability to learn important features.

[0009] Preferably, the formula of the synergistic effect function is: Among them, G(X i ,X j ) is feature X i and X j The value of the synergistic effect function between i ) is feature X i Conditional entropy of the target variable Y; H(Y|X j ) is feature X j The conditional entropy of the target variable Y; ∈ is a positive smoothing factor; X i and X j is a feature in member behavior data, time series data, product data or member feature data; Y is the target churn label; The weight distribution satisfies the constraints: Among them, w is the feature X i The weight of ; n is the number of features after screening.

[0010] Preferably, the time series model is a long short-term memory network, and its network structure includes: Input layer: receives the optimized feature set 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 value of member churn probability through the fully connected layer.

[0011] Preferably, the input features of the time series model are calculated by weighting: The input feature vector is represented as: in, For feature X i The weight of X is the i-th eigenvalue at the t-th time point in the time series; t,1 is the first feature value at the tth time point in the time series; n is the total number of input features, X t is the weighted input feature vector.

[0012] Preferably, the evaluation model performance index comprises the following steps: Calculate the accuracy, recall, F1 score, and area under the curve of the prediction results; Recalculate feature weights and adjust model parameters based on the entropy value extracted from the time series; Update the model regularly to ensure its robustness under different data distribution conditions.

[0013] Preferably, the present invention also provides a member churn prediction system based on model evaluation, comprising the following modules: Data processing module: used to clean, normalize and extract time series features from member behavior data; Feature optimization module: used to calculate information entropy and conditional entropy and screen key features; Weight allocation module: dynamically optimize feature weights based on game theory; Time series prediction module: Based on the long short-term memory network model, train and predict the probability of member churn; Model evaluation module: evaluates prediction results and dynamically optimizes model parameters.

[0014] The present invention provides a member churn prediction method based on model evaluation. It has the following beneficial effects: 1. The present invention adopts a feature optimization method that combines information entropy and conditional entropy. By quantifying the explanatory power of each feature on the loss label, the key features are screened out, achieving the technical effect of improving model training efficiency and reducing interference from irrelevant features. Compared with the technical solution in the prior art that only relies on manual experience or simple statistical methods for feature selection, it solves the problem of model performance degradation due to insufficient feature importance evaluation.

[0015] 2. The present invention adopts a dynamic feature weight allocation mechanism based on game theory. By constructing a feature synergy effect function and optimizing weight allocation, it maximizes the contribution of key features to model learning and achieves the technical effect of optimizing model learning ability. Compared with the technical solution in the prior art where feature weights are fixed or not dynamically adjusted, it solves the problem that the model accuracy decreases due to changes in feature importance over time and data distribution.

[0016] 3. The present invention uses long short-term memory network to model time series data, combined with weighted optimization features, to capture the time dependency in member behavior data, and achieves the technical effect of accurately predicting member churn trends. Compared with the technical solution in the prior art where traditional machine learning models are difficult to handle time dependency, this solves the problem of low accuracy in predicting member churn and insufficient ability to learn time series features.

[0017] 4. The present invention adopts a dynamic evaluation and optimization mechanism, evaluates the model prediction effect through performance indicators, and dynamically adjusts the feature weights and model parameters in combination with the time series entropy value increase, thereby achieving the technical effect of enhancing the adaptability and robustness of the model. Compared with the static model training method in the prior art, it solves the shortcomings of the model being insensitive to changes in data distribution and unstable performance in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention; Figure 2 is a system architecture diagram of the present invention; DETAILED DESCRIPTION

[0019] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] Please see attached Figure 1 The present invention provides a member churn prediction method based on model evaluation, which realizes accurate prediction of member churn trend through a multi-step technical combination, including extracting and cleaning data from the member management system, optimizing and screening features based on information entropy, constructing feature synergy effect function and dynamically allocating weights using game theory, combining time series model to train churn prediction model and dynamic evaluation optimization strategy.

[0021] It should be noted that the high-quality processing of the data in step S1 provides input support for the feature screening in S2. The screening results of S2 further optimize the feature weights through S3. The weights optimized by S3 are directly used for the time series model training in S4, while S5 reacts to the above steps through the evaluation and optimization mechanism.

[0022] like Figure 1 As shown, a member churn prediction method based on model evaluation of the present invention may include the following steps: S1. Data extraction and preprocessing; S2. Feature optimization and screening; S3. Feature weight assignment; S4. Training of time series models; S5. Model evaluation and optimization.

[0023] For step S1, in this embodiment, it is divided into the following four steps: Data extraction Member behavior data: including page visit frequency, product clicks, number of times products are added to shopping carts, number of completed transactions, average transaction amount, etc. For example, a member's visit records in the past three months show that he clicked on a certain product page 10 times and completed 3 purchases. These behavior data are automatically recorded by the user behavior log system; Time series data: including the date and time of specific operations, the time interval between behaviors, and the weekly or monthly active behavior statistics of members. As an option, time series data can also extract daily transaction frequency or behavior patterns to capture dynamic changes; Member feature data: including static features such as member registration time, level, region, and dynamic features such as loyalty score and last visit time; Churn label data: generated by rules with no active behavior or historical annotations in the last three months, defined as a binary label: Y=1 indicates churn, and Y=0 indicates no churn; in some embodiments, more complex label rules can be used, such as defining labels in combination with key business indicators.

[0024] Data cleaning Missing value processing: For numerical features such as consumption amount and visit frequency, the mean filling method is used; for categorical features such as member level and product category, the mode filling method is used; for time series data, linear interpolation is used. For example, if the visit frequency record of a member in a certain month is missing, the average of the visit frequency in the last three months can be used instead.

[0025] Redundant feature removal: Remove features that are irrelevant or have low relevance to member churn prediction, such as fixed member IDs that are completely static and have no temporal significance.

[0026] Normalization In order to ensure the dimensional consistency between different features, all numerical data are normalized. The formula is: It can be understood that normalization effectively avoids model instability caused by eigenvalue range being too large or too small.

[0027] Among them, x is the original eigenvalue, min(x) is the minimum value of feature x in the entire data set, max(x) is the maximum value of feature x in the entire data set, and x ′ is the normalized eigenvalue.

[0028] Temporal feature extraction Behavior time interval: Calculate the time interval between consecutive behaviors using timestamps: Δt=t i+1 -t i For example, if two consecutive purchase records are made on January 1, 2024 and January 10, 2024, the time interval is 9 days; Time window statistics: Statistics on the behavior characteristics of a fixed time window (such as the last 30 days). For example, calculate the average consumption amount and transaction frequency of members in the last month; Where: t i The timestamp of the i-th behavior is used to extract periodic behavior features, such as the total monthly consumption amount and the number of weekly visits; and the user's behavior active stage (initial registration active period and stable period) is analyzed.

[0029] For step S2, in this embodiment, the importance of each feature is quantitatively evaluated by information entropy and conditional entropy, redundant features are eliminated, and key features that contribute significantly to churn prediction are retained. Information entropy is used to measure the distribution complexity of a single feature. It is defined as the total amount of information uncertainty in the probability distribution of feature values.

[0030] For example: if a feature has a uniform distribution of values ​​in all samples, the entropy of the feature is high; on the contrary, if a feature is highly concentrated in certain specific values, the entropy is low.

[0031] Conditional entropy is used to evaluate the ability of features to explain whether the target label has been lost or not, that is, the uncertainty of the target label under the condition that the feature is known. Features with lower conditional entropy indicate that they have stronger explanatory power for the target. By combining information entropy and conditional entropy, by setting the conditional entropy threshold, features that meet the threshold requirements are screened. Features with conditional entropy lower than the set threshold are considered to have a strong contribution to member churn prediction. In addition, in order to dynamically capture member behavior trends, the entropy value of the time interval is calculated for the time series features, and dynamic features with significant behavior changes are extracted from them.

[0032] Information entropy measures the complexity of feature data distribution, and the calculation formula is: In some embodiments, the calculation of information entropy is applicable to discrete features, such as membership level, product category, etc. It should be noted that for consumption amount and access frequency, they can be discretized by bucketing and then the information entropy can be calculated; Conditional entropy is used to quantify the explanatory power of features for the churn labels: In one possible implementation, features with smaller conditional entropy values ​​have stronger explanatory power for churn labels and are therefore more valuable for screening; Feature screening rule: set conditional entropy threshold δ H , filter the features that meet the following conditions: As an option, the threshold δ H The setting can be adjusted according to business needs. For example, when the number of model input features needs to be reduced, a stricter threshold can be selected to filter out the most influential features; It should be noted that feature screening is not limited to conditional entropy, but can also be combined with the calculation results of information entropy. For example, features with high information entropy but low conditional entropy can be retained first, thereby ensuring that the input features have rich distribution information and are highly relevant to the churn label; Time series feature optimization: Calculate the entropy value of time series data: ΔH t =H t -H t-1 Extract features with significant dynamic changes. In one possible implementation, time features with larger absolute values ​​of entropy increment can be retained first, because these features often represent significant changes in the data. Among them, H(Y|X) is the conditional entropy; P(x i ,y j ) is the value x of feature X i and the value y of the target variable Y j The joint probability of j ∣x i ) is the value of feature X when i Under the condition that j The conditional probability of ; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y; is the set of features with significant dynamic changes; ΔH t is the entropy increment of the time series feature at time t; δ H Entropy value threshold for dynamic feature screening.

[0033] For step S3, in this embodiment, a weight optimization method based on game theory is adopted to dynamically assign weights to the screened feature set, so that the model can better focus on key features. The feature synergy effect 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 target labels, the complementary effects between different features are evaluated. By optimizing the total effect of feature weight allocation, ensure that the sum of weights is 1 and the weights are non-negative. The purpose of dynamically optimizing weights is to highlight key features while weakening the influence of redundant features. An iterative optimization algorithm (gradient descent) is used to calculate feature weights so that they gradually converge under the goal of optimal weight allocation. The dynamically adjusted weights are directly used as input weighting factors for subsequent time series models; Construction of synergy effect function, feature X i and Xj The synergy function is defined as: 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, and the model tends to assign more weight to features with lower conditional entropy; The optimization goal of weight distribution is to maximize the total contribution of synergy: Among them, w i The feature w j In a possible implementation, the weight distribution must satisfy the following constraints: Iterative optimization, using gradient descent to update weights: Among them, X i and X j is a feature of member behavior data, time series data, product data or member feature data; Y is the target churn label; G(X i ,X j ) is feature X i and X j The value of the synergistic effect function between i ) is feature X i Conditional entropy of the target variable Y; H(Y|X j ) is 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 weight of ; η is the learning rate.

[0034] For step S4, in this embodiment, the weighted input of the time series model is formed after the filtered feature set is weighted and optimized. The input features are weighted according to the weights, which helps the model to more accurately capture the contribution of key features to churn prediction. The long short-term memory network is used as the core structure of the time series model. The time series model is used to capture the dynamic characteristics of member behavior over time. Combined with the filtered key features and their weights, the prediction model is trained to generate the probability output of member churn: Input layer: receives weighted feature input.

[0035] Hidden layer: Capture the long-term and short-term dependencies of time series through long short-term memory network units.

[0036] Output layer: The hidden layer state is mapped to the probability of member churn through the fully connected layer.

[0037] The loss function combines the prediction error and feature weight optimization objectives, while considering the deviation between the model prediction results and the true labels, as well as the regularization constraints of the feature weights on the model. This design ensures that key features are learned first while controlling the complexity of the model; The loss function combines cross entropy and entropy regularization: Among them, L is the total loss function value; N is the total number of training samples; y i is the true label of the i-th sample; is the predicted churn probability of the model for the i-th sample; and represents the logarithm of the predicted probability; λ is the regularization parameter; w j For feature X j The weight of H(Y|X j ) is feature X j The conditional entropy of; n is the total number of features after screening; In an exemplary implementation, after training, the model generates a prediction of the probability of churn for the input at each time point. To evaluate the performance of the model, the following indicators can be used: accuracy: measures the proportion of correct predictions made by the model, recall: measures the model's ability to detect churned members, F1 score combines precision and recall as a comprehensive evaluation indicator, and area under the curve: reflects the model's ability to distinguish between churned and non-churned members.

[0038] For step S5, in this embodiment, the performance of the prediction model is evaluated, and combined with the actual loss label comparison analysis, the feature weights and model parameters are dynamically optimized to improve the overall prediction effect of the model. It should be noted that this step is closely related to the previous steps in logic and is an important link for ensuring the robustness of the model in different data distributions and scenarios.

[0039] As an implementation, performance metrics are calculated based on the churn probability predicted by the model and the actual churn label.

[0040] If it is found that the weight distribution of some features is insufficient, the weights are redistributed through entropy appreciation, and the optimized features are adjusted and input into the time series model.

[0041] Depending 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.

[0042] After the final optimized model, the evaluation process is repeated to ensure that the adjusted model has better prediction performance under different data distributions.

[0043] Please see attached Figure 2 The present invention also provides a member churn prediction system based on model evaluation, which combines data processing, feature optimization, weight allocation, time series modeling and dynamic optimization technology to ensure the accuracy and efficiency of member churn prediction.

[0044] The member churn prediction system based on model evaluation in this embodiment mainly includes the following functional modules: data processing module, feature optimization module, weight allocation module, time series prediction module, model evaluation and optimization module. Each module collaborates to form a closed-loop system to continuously improve the prediction effect.

[0045] Data processing module: Responsible for extracting and processing raw data from the member management system to provide high-quality input for subsequent modules. The main functions include: data extraction: reading member behavior data, time series data, member feature data and churn label data from the database; data cleaning: clearing invalid data, filling missing values, and removing outliers; data standardization: normalizing numerical features and unifying data dimensions; time feature extraction: generating time interval features and periodic behavior features.

[0046] Feature Optimization Module: Based on information entropy and conditional entropy, the data features are quantitatively evaluated and screened, redundant features are eliminated, and key features are extracted. Implementation process: Calculate the information entropy H(X) of each feature to measure the complexity of data distribution; calculate the conditional entropy H(Y|X) of the feature to evaluate the feature's ability to explain the loss label; set a threshold based on the conditional entropy and screen the features that meet H(Y|X i )≤δ H The key features of the time series are calculated by entropy increment and features with significant dynamic changes are extracted.

[0047] Weight distribution module Use the game theory optimization model to assign weights to the selected features and dynamically adjust the importance of the features. Implementation process: Construct the synergy effect function G(X i ,X j ), measure the correlation between different features; define the optimization objective function of weight distribution, and solve the feature weights through gradient descent method; dynamically update the weights to ensure that key features contribute more to model training.

[0048] Time Series Forecasting Module: Based on the weighted feature set, the time series model is used to predict the probability of member churn. Main functions: Receive the weighted optimized feature set as input; build a time series model, including the input layer, hidden layer time series units and output layer; use the weighted input features to train the time series model and output the probability of member churn.

[0049] Model evaluation and optimization module: The performance of the prediction model is evaluated through multiple evaluation indicators, and the model parameters and feature weights are adjusted dynamically. Implementation process: Use indicators such as accuracy, recall, F1 score and AUC to evaluate model performance; dynamically adjust feature weights according to entropy value to optimize model input; regularly update the parameters of the time series model to improve the robustness of the model under different data distributions.

[0050] The system module of this embodiment is designed based on the execution process of the above method embodiment. Its principles and technical effects are similar to those of the method embodiment and will not be repeated here.

[0051] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A member churn prediction method based on model evaluation, characterized in that: The following steps are involved: Extract member behavior data, time series data, member feature data, and churn label data from the member management system, perform data cleaning and normalization, and extract time features; 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 the preset threshold are screened; Construct the synergy effect function between features and determine the dynamic weight distribution scheme of features through the game theory model; Based on the screened features and corresponding weights, the member churn prediction model is trained using a time series model to generate prediction outputs. By comparing the prediction model output with the actual churn labels, the model performance indicators are evaluated, and the feature weights and model parameters are dynamically adjusted according to the entropy value.

2. A 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 feature data and churn label data from the member management system, performing data cleaning and normalization processing, and extracting time features include: Extracting time interval features and periodic behavior features from time series data, including calculating the time interval change rate of member behavior and extracting weekly and monthly periodic features of behavior data; The numerical features are normalized and the minimum and maximum normalization methods are used to map the feature values ​​to the [0,1] interval to unify the data scale.

3. A member churn prediction method based on model evaluation according to claim 1, characterized in that: The information entropy and conditional entropy calculation is used to calculate the information entropy of each feature, measure the complexity of the feature and the uniformity of data distribution; the explanatory power of each feature for member churn prediction is calculated based on the conditional entropy, and the feature with a smaller conditional entropy value has a greater contribution to the churn label; The features whose entropy values ​​are lower than the preset threshold are selected as key features; For time series data, the features with significant dynamic change trends are extracted by calculating the entropy value of time features.

4. A member churn prediction method based on model evaluation according to claim 3, characterized in that: The information entropy is defined as: The conditional entropy is defined as: Among them, H(Y|X) is the conditional entropy; P(x i ,y j ) is the value x of feature X i and the value y of the target variable Y j The joint probability of j ∣x i ) is the value of feature X when i Under the condition that j The conditional probability of ; n is the number of possible values ​​or segments of feature X; m is the number of possible values ​​of target variable Y.

5. A member churn prediction method based on model evaluation according to claim 1, characterized in that: The synergistic effect function between the constructed features is used to quantify the correlation between different features; The game theory model is used to take feature weight allocation as the optimization goal, 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.

6. A member churn prediction method based on model evaluation according to claim 5, characterized in that: The formula of the synergistic effect function is: Among them, G(X i ,X j ) is feature X i and X j The value of the synergistic effect function between i ) is feature X i Conditional entropy of the target variable Y; H(Y|X j ) is feature X j The conditional entropy of the target variable Y; ∈ is a positive smoothing factor; X i and X j is a feature in member behavior data, time series data, product data or member feature data; Y is the target churn label; The weight distribution satisfies the constraints: Among them, w is the feature X i The weight of ; n is the number of features after screening.

7. A 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 optimized feature set 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 value of member churn probability through the fully connected layer.

8. A member churn prediction method based on model evaluation according to claim 7, characterized in that: The input features of the time series model are calculated according to the weights: The input feature vector is represented as: in, For feature X i The weight of X is the i-th eigenvalue at the t-th time point in the time series; t,1 is the first feature value at the tth time point in the time series; n is the total number of input features, X t is the weighted input feature vector.

9. A member churn prediction method based on model evaluation according to claim 1, characterized in that: The evaluation model performance index comprises the following steps: Calculate the accuracy, recall, F1 score, and area under the curve of the prediction results; Recalculate feature weights and adjust model parameters based on the entropy value extracted from the time series; Update the model regularly to ensure its robustness under different data distribution conditions.

10. A member churn prediction system based on model evaluation, applied to a member churn prediction method based on model evaluation as described in claims 1-9, characterized in that: The member churn prediction system based on model evaluation includes the following modules: Data processing module: used to clean, normalize and extract time series features from member behavior data; Feature optimization module: used to calculate information entropy and conditional entropy and screen key features; Weight allocation module: dynamically optimize feature weights based on game theory; Time series prediction module: Based on the long short-term memory network model, train and predict the probability of member churn; Model evaluation module: evaluates prediction results and dynamically optimizes model parameters.

Images