A prediction method for user churn in logistics enterprises

Through the combination of CDF cumulative distribution function and RF and CRITIC methods, and the Tianniu Xu search algorithm combined with BAS and SVM, the accuracy of user churn prediction of logistics enterprises is solved, and more efficient user churn prediction and risk assessment are achieved, which enhances the Competitiveness of the enterprise.

CN114677167BActive Publication Date: 2024-12-06NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210236265.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-11
Publication Date
2024-12-06
Estimated Expiration
2042-03-11

AI Technical Summary

Technical Problem

Logistics companies have challenges in predicting user churn, and existing technologies are difficult to accurately predict user churn, resulting in reduced benefits and increased competition.

Method used

Data conversion is performed using CDF cumulative distribution function, combined with random forest (RF) and CRITIC methods for feature extraction and user churn coefficient evaluation, and through the combination of BAS and SVM, the Tianniu Xuxu search algorithm is used to optimize the penalty factor and kernel function parameters. , achieve more accurate user churn prediction.

Benefits of technology

It improves the accuracy and efficiency of user churn prediction, reduces the risk of user churn, and enhances the competitiveness of logistics companies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114677167B_ABST
    Figure CN114677167B_ABST
Patent Text Reader

Abstract

The present invention provides a method for predicting user churn of a logistics enterprise, comprising the following steps: S1: preprocessing the original data set of users of the logistics enterprise, using the CDF cumulative distribution function to convert the data, and making the data dimensionless; S2: feature engineering stage processing: using the RF and CRITIC methods to extract features from the data set and evaluate the user churn coefficient; S3: predicting and processing user churn, using a method combining BAS and SVM to make the local and global optimal extreme values ​​more accurate. The present invention can realize the prediction of user churn by logistics enterprises and prevent user churn.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for predicting user churn of a logistics enterprise, and belongs to the technical field related to logistics and machine learning. Background Art

[0002] With the rapid development of the logistics industry, the number of users of logistics companies is increasing, including many franchisees and some small trading users. With the development of logistics companies and more and more users joining, the competition among users is gradually expanding. The reduction of benefits and the increase of competition have led some users to choose to leave logistics companies. Therefore, the prediction of user churn is very important for logistics companies.

[0003] This patent designs a method for predicting user churn for logistics companies. The method includes the following algorithms, which are all prior art. This patent involves the application of the following algorithms.

[0004] Cumulative distribution of data: that is, the probability distribution of all data less than or equal to the current data value, which is very helpful in expressing the probability of a data point appearing in a certain interval. Mathematically speaking, the cumulative distribution function (Cumulative Distribution Function) is the integral of the probability distribution function. The cumulative distribution function has the following characteristics: 1. The cumulative distribution function is a monotonically increasing function on the x-axis. 2. The cumulative distribution function is smoother and has less noise in the image. 3. The cumulative distribution function does not introduce external concepts such as bandwidth, so no data information will be lost. For a given data set, the cumulative distribution function is unique. 4. The cumulative distribution function is generally normalized, monotonically increasing and close to 1.

[0005] In machine learning, there is often a need to convert data of different specifications to the same specification, or to convert data of different distributions to a specific distribution. This need needs to be solved by dimensionless data. Linear dimensionless includes centralization and scaling. The essence of centralization is to subtract a fixed value from all records, that is, to shift the data sample data to a certain position. The essence of scaling is to fix the data within a certain range by dividing it by a fixed value. Taking the logarithm is also a kind of scaling.

[0006] Random forest is a random forest that is built in a random way. There are many decision trees in the forest. There is no correlation between each decision tree in the random forest. After the forest is obtained, when a new input sample enters, each decision tree in the forest will make a judgment to see which category the sample should belong to, and then see which category is selected the most, and then predict the sample to that category. Random forest can handle both discrete and continuous values, and can also be used for unsupervised learning clustering and outlier detection.

[0007] A decision tree is a tree structure in which each non-leaf node represents a test on a feature attribute, each branch represents the output of this feature attribute in a certain value range, and each leaf node stores a category. The process of using a decision tree to make decisions is to start from the root node, test the corresponding feature attributes in the item to be classified, and select the output branch according to its value until the leaf node is reached, and the category stored in the leaf node is used as the decision result.

[0008] The CRITIC method is a better objective weighting method than the entropy weight method and the standard deviation method. It comprehensively measures the objective weight of the indicators based on the comparative strength of the evaluation indicators and the conflict between the indicators. It takes into account the variability of the indicators while taking into account the correlation between the indicators. The larger the number, the more important it is. It fully utilizes the objective attributes of the data itself for scientific evaluation.

[0009] The contrast intensity refers to the size of the difference between the values ​​of the various evaluation schemes for the same indicator, which is expressed in the form of standard deviation. The larger the standard deviation, the greater the fluctuation, that is, the greater the difference between the values ​​of the schemes, the higher the weight will be; the conflict between indicators is expressed by the correlation coefficient. If there is a strong positive correlation between two indicators, it means that the conflict is smaller and the weight will be lower.

[0010] The beetle whisker search algorithm is a biologically inspired intelligent optimization algorithm. It is an algorithm developed based on the foraging principle of beetles. Its bionic principle is as follows:

[0011] When the longhorn beetle forages for food, it does not know where the food is, but forages based on the strength of the food's smell. The longhorn beetle has two long antennae. If the left antennae receive a stronger smell than the right antennae, the next step is to fly to the left, otherwise it will fly to the right. Based on this principle, the longhorn beetle can find food. The smell of food is equivalent to a function, and the value of each point in this function is different in three-dimensional space. The longhorn beetle's two antennae can collect the smell values ​​of two points near itself. The longhorn beetle's goal is to find the point with the largest global smell value (that is, the location of the food). Based on the behavior of the longhorn beetle, we designed this intelligent optimization algorithm to solve the function optimization.

[0012] Support vector machines (SVM) are a binary classification model. Its basic model is a linear classifier with the largest margin defined in the feature space. The largest margin makes it different from the perceptron. SVM also includes kernel techniques, which makes it a nonlinear classifier in essence. The learning strategy of SVM is to maximize the margin, which can be formalized as a problem of solving a convex quadratic programming, which is equivalent to the problem of minimizing the regularized hinge loss function. Summary of the invention

[0013] The purpose of the present invention is to provide a method for predicting user churn of a logistics enterprise, which is used to predict the user churn of a logistics enterprise.

[0014] The technical solution of the present invention is as follows: 1. A method for predicting user churn of a logistics enterprise, comprising the following steps:

[0015] S1: Data preprocessing is performed on the original data set of logistics enterprise users. For the original data set, CDF cumulative distribution function is used for data conversion, and all features are graded and classified to make all data dimensionless;

[0016] S2: Feature engineering stage processing: RF and CRITIC methods are used to extract data features and evaluate user churn coefficients;

[0017] S3: The prediction and processing of user churn adopts a method combining BAS and SVM, combining the idea of ​​population with the BAS method, and adding the idea of ​​population and optimal solution guidance mechanism into BAS. Through BAS, the penalty factor and kernel function parameters of SVM are better optimized, making the local and global optimal extreme values ​​more accurate.

[0018] Furthermore, in the above-mentioned method for predicting user churn of logistics enterprises, wherein: each feature in the original data set in step S1 is classified and processed, and the features include basic information data, user behavior attribute data and logistics enterprise attribute data, the basic information includes user number, user identity, user scale, and market, the user behavior attribute data includes service status, transaction quantity, income and expenditure status, the logistics enterprise attribute data includes transaction status and financial subsidies, and each feature is field named and coded.

[0019] Furthermore, the above-mentioned method for predicting user churn of logistics enterprises, wherein: using the CDF cumulative distribution function method to perform data conversion, the specific method diagram is as follows:

[0020] For the features after hierarchical processing, CDF cumulative distribution function is used to convert the data of each feature. All data in each feature are expressed as x(x1,x2,x3,...,xn ), where n is the total data of each indicator, and the sorting order of each value in the indicator is expressed as (x [1] ,x [2] ,x [3] ,...,x [n] ), the conversion process of indicator data is as follows:

[0021]

[0022] For negative indexes, the opposite index number is first used, and then the CDF cumulative distribution function is used to transform the data.

[0023] Furthermore, in the above method for predicting user churn of logistics enterprises, step S2 specifically includes the following steps:

[0024] S21: Extract and screen the feature indicators in the data set through RF. RF is composed of multiple decision trees. The generation of each decision tree follows the top-down division principle. The training set is divided from the root node for each decision tree. For each node in the tree, its division attribute is selected based on the principle of minimum purity node, and the training data set is separated at the same time.

[0025] S22: Use the CRITIC method to assess the risk of user churn.

[0026] Furthermore, in the above-mentioned method for predicting user churn of logistics enterprises, in which: in the step S21, the decision tree generates child nodes during training, and when the decision tree meets the branch stop rule, the growth of the child nodes is stopped. In order to determine the most appropriate feature attribute, it is necessary to quantify the "purity" of each feature, and the method for calculating the "purity" of each feature is as follows:

[0027] S211: Calculate the gain of information entropy in the feature:

[0028]

[0029] Where S is the sample set for training, F is the feature contained in the sample, P(F|i) is the probability of each category when the feature appears, and P(i) is the probability of the feature appearing in the sample;

[0030] S212: Use the features in the sample as random variables and divide the sample S for training:

[0031]

[0032] S213: Calculate the "purity" of the features in the sample set and extract the features with high "purity":

[0033]

[0034] Furthermore, in the above method for predicting user churn of logistics enterprises, step S22 includes the following steps:

[0035] S221: Calculate the conflict quantitative index between the jth characteristic index and other indexes:

[0036]

[0037] Among them, the correlation coefficient The covariance and standard deviation of the samples are selected for representation. S222: Calculate the amount of information contained in the j-th feature:

[0038]

[0039] S223: Calculate the objective weight of the jth feature:

[0040]

[0041] X ij represents the jth secondary indicator in the i-th primary indicator, Represents the average value of the i-th first-level indicator.

[0042] Furthermore, in the above method for predicting user churn of logistics enterprises, step S3 specifically includes the following steps:

[0043] S31: Obtain the original data set of enterprise users, set the population size n, and the maximum number of iterations Ite max , set the value range of the optimal parameter set (C (penalty factor), g (kernel function)), divide the data set into two parts, the training set and the test set;

[0044] S32: Initialize the position of the longhorn beetle group, optimize the penalty factor (C) and kernel parameter (g) of SVM, and the position of each longhorn beetle represents a pair of parameters (C, g);

[0045] S33: Determine the position coordinates of the left and right whiskers of the longicorn:

[0046]

[0047]

[0048] Where t is the number of loops of the method, ran(k,1) is a random k-dimensional vector, and distance represents the distance between the left and right whiskers of the longicorn.

[0049] S34: Determine the fitness function. The fitness function selected by this method is as follows:

[0050]

[0051] Among them, W i is the current weight factor, X i To obtain real data, To fit the data, calculate the fitness value of each individual beetle and record the optimal value of the current individual and population;

[0052] S35: Start iteration and update the population position;

[0053] S36: Calculate the fitness value after the position is updated, update the individual position by comparing the fitness values, and obtain a new population. Determine whether to continue updating or output the optimal parameter pair based on whether the preset maximum number of iterations is reached or the preset fitness function accuracy is reached. The iteration ends when the maximum number of iterations is reached or the preset fitness function accuracy is reached.

[0054] S37: Obtain an optimal parameter pair (C, g), use the optimal parameters to train and model the training set samples, and obtain an output optimal parameter combination;

[0055] S38: Substitute the test set data into the training model after step S38 for testing to obtain the SVM prediction model.

[0056] Furthermore, in the above method for predicting user churn of logistics enterprises, step S35 specifically includes the following steps:

[0057] Optimize the step size parameter in the position update formula and change the original static step size coefficient to a dynamic adjustment coefficient.

[0058]

[0059]

[0060] in, is the maximum step length, Ite is the current number of iterations, Ite max is the maximum number of iterations. This method can effectively step Restricted to Within,

[0061] Update the position of the longhorn beetle and introduce the MPSO typical optimal solution guidance mechanism to achieve the goal of making the longhorn beetle group approach the optimal individual solution faster. Define the position formula of the longhorn beetle:

[0062]

[0063] Among them, Cons is a constant defined between [0,1], Locbest is the position of the optimal individual under the current number of iterations, ran(k,1) is a random k-dimensional vector, and sign(.) is the sign function:

[0064]

[0065] The present invention performs data preprocessing on the data set of the entire logistics enterprise user, converts the data through the CDF method, and uses the cumulative distribution function to perform dimensionless operations on the data. In the feature engineering stage after preprocessing, the RF and CRITIC methods are used to screen the data and evaluate the user churn coefficient. In the prediction method of user churn, a method combining BAS and SVM is adopted, and the idea of ​​the population and the optimal solution guidance mechanism is added to BAS to accelerate the speed of the longhorn beetle population approaching the optimal individual solution. The penalty factor and kernel function parameters of SVM are better optimized through BAS, which is used to predict user churn and prevent user churn. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 It is a flow chart of the user churn prediction method;

[0067] Figure 2 It is the feature classification map of the original data set. DETAILED DESCRIPTION

[0068] In order to make the implementation objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention are described clearly and completely in steps below in conjunction with the accompanying drawings.

[0069] like Figure 1 , is a prediction method for user churn, including the following steps:

[0070] S1: Perform data preprocessing on the original data set of logistics enterprise users. For the original data set, use the CDF cumulative distribution function to convert the data, classify all features, and make all data dimensionless. In this step, for the original data set initially obtained, there must be some duplicate, abnormal, or even missing data in the data. This part of the data needs to be processed before the feature classification and extraction project to minimize the impact on subsequent modeling. For some data in the original data set that is not strongly correlated with the predicted user churn, these data need to be deleted.

[0071] S2: Feature engineering stage processing: RF and CRITIC methods are used to extract data features and evaluate the user churn coefficient. This step is an engineering activity used to convert data into a feature system that can be understood by computer algorithms, improving the accuracy and generalization ability of the model. For the data after data conversion, its features are classified, but because there are too many variables directly used to predict user churn indicators, the indicators are too complicated, so fewer and more representative indicators are selected to measure user churn. The indicators in the indicator system are continuous variables and can be screened through machine learning.

[0072] S3: User churn prediction and processing, using a combination of BAS and SVM, combining the idea of ​​population with the BAS method, introducing the MPSO typical optimal solution guidance mechanism into BSA to speed up the speed of the beetle group approaching the optimal individual solution, and using BAS to better optimize the penalty factor and kernel function parameters of SVM, so that the local and global optimal extreme values ​​are more accurate. In this step, since credit data is complex, with many features and both continuous and discrete types, selecting effective features from high-dimensional data plays a vital role in the final credit assessment results.

[0073] Step S1 includes the following steps: some enterprise and user indicators are selected. There are too many variables directly used to predict user churn indicators, which makes the indicators too complicated. Therefore, fewer and more representative indicators are selected to measure user churn. The indicators in the indicator system are continuous variables and can be screened through machine learning. The classification processing of each enterprise and user indicator in the original data set is as follows: Figure 2 As shown, the index system includes basic information data, user behavior attribute data and logistics enterprise attribute data. The basic information includes user number, user identity, user scale, market, etc. The user behavior attribute data includes service status, transaction quantity, income and expenditure status, etc. The logistics enterprise attribute data includes transaction status, financial subsidies, etc.

[0074] All indicators in the above original data set are graded and classified so that they can better cooperate with the CDF cumulative distribution function to make the data dimensionless. The CDF cumulative distribution function method is used for data conversion. Compared with general normalization and standardization algorithms, it is more convenient and effective. It does not require samples to obey normal distribution, and it maximizes the preservation of the true characteristics of the original data and reduces the sensitivity of extreme values ​​to new samples. Its usage is as follows:

[0075] For the indicators after hierarchical processing, CDF cumulative distribution function is used to convert the data of each indicator. Figure 2 All data in each indicator are represented as x(x1,x2,x3,...,x n), where: r represents the current indicator, n is the total data of each indicator. The sorting order of each value of the indicator is expressed as (x [1] ,x [2] ,x [3] ,...,x [n] ), the conversion process of indicator data is as follows:

[0076]

[0077] In step S2, feature engineering module processing is performed to classify the features of the data after data conversion, which specifically includes the following steps:

[0078] S21: After preprocessing the original data set, RF is used to extract and screen the feature indicators in the data set. RF is composed of multiple decision trees. The generation of each decision tree follows the top-down division principle, so the method of dividing the training set from the root node is adopted for each decision tree. For each node in the tree, based on the principle of minimum purity node, its division attribute is selected, and the training data set is separated at the same time.

[0079] The decision tree will generate child nodes during training. When the decision tree meets the branch stopping rule, the growth of child nodes will stop. In order to determine the most appropriate feature attributes, it is necessary to quantify the "purity" of each feature. The "purity" of the feature can be quantified by calculating the information gain ratio of the feature index. The higher the information gain ratio, the higher the "purity" of the feature. After feature F is trained, the sample S is divided into multiple categories. The method for calculating the "purity" of each feature is as follows:

[0080] Step 1: Calculate the gain in information entropy in the feature:

[0081]

[0082] Among them, S is the sample set for training, F is the feature contained in the sample, P(F|i) is the probability of each category when the feature appears, and P(i) is the probability of the feature appearing in the sample.

[0083] Step 2: Use the features in the sample as random variables and divide the sample S for training:

[0084]

[0085] Step 3: Calculate the "purity" of the features in the sample set and extract the features with high "purity":

[0086]

[0087] S22: Use the CRITIC method to assess the risk of user churn: The risk index of user churn in logistics enterprises is synthesized based on the characteristic indicators of the user data set. Compared with using historical data for regression modeling to monitor user churn risk, the comprehensive index method is more convenient and flexible, and can be combined with other model methods for research. Specifically, it includes the following steps:

[0088] Step 1: Calculate the conflict quantitative index between the jth feature index and other indicators:

[0089]

[0090] Among them, the correlation coefficient The covariance and standard deviation of the samples are selected for representation, X ij represents the jth secondary indicator in the i-th primary indicator, Represents the average value of the i-th first-level indicator.

[0091] Step 2: Calculate the amount of information contained in the jth feature:

[0092]

[0093] Step 3: Calculate the objective weight of the jth feature:

[0094]

[0095] In step S3, the user churn prediction module uses a method combining BAS and SVM to predict user churn. In order to obtain the optimal parameters, the MPSO algorithm is combined with the BAS method, and the MPSO typical optimal solution guidance mechanism is introduced to make the longhorn beetle group approach the optimal individual solution faster. Specifically, the following steps are included:

[0096] S31: Obtain the original data set of enterprise users, set the population size n, and the maximum number of iterations Ite max , set the value range of the optimal parameter set (C (penalty factor), g (kernel function)), divide the data set into a training set and a test set. The training set is used to select the SVM model with better performance; the test set is used to test the performance of the optimized SVM model.

[0097] S32: Initialize the position of the longhorn beetle group. Since the parameters that need to be optimized are the penalty factor (C) and kernel parameter (g) of the SVM, the space for searching the longhorn beetle group is set to two dimensions, and the position of each longhorn beetle represents a pair of parameters (C, g).

[0098] S33: Determine the position coordinates of the left and right whiskers of the longicorn:

[0099]

[0100]

[0101] Where t is the number of loops of the method, ran(k,1) is a random k-dimensional vector, and distance represents the distance between the left and right whiskers of the longicorn.

[0102] S34: Determine the fitness function. The fitness function selected by this method is as follows:

[0103]

[0104] Among them, W i is the current weight factor, X i To obtain real data, The data obtained for fitting is used to calculate the fitness value of each individual beetle and record the optimal value of the current individual and population.

[0105] S35: Start iteration and update the population position. The step length parameter in the position update formula is optimized as follows:

[0106] Inspired by the optimization of inertia weight in the population concept, a linear decreasing strategy based on the current number of iterations is proposed to optimize the step coefficient of the current number of iterations. Changing the original static step coefficient to a dynamic adjustment coefficient helps to optimize the convergence speed of the BAS algorithm and improve the diversity of solutions to a certain extent.

[0107]

[0108]

[0109] in, is the maximum step length, Ite is the current number of iterations, Ite max is the maximum number of iterations. This method can effectively step Restricted to Within.

[0110] Update the position of the longhorn beetle and introduce the MPSO typical optimal solution guidance mechanism to achieve the goal of making the longhorn beetle group approach the optimal individual solution faster. Define the position formula of the longhorn beetle:

[0111]

[0112] Among them, Cons is a constant defined between [0,1], Loc best is the position of the optimal individual under the current number of iterations, ran(k,1) is a random k-dimensional vector, and sign(.) is the sign function:

[0113]

[0114] S36: Calculate the fitness value after the position is updated, update the individual position by comparing the fitness value, and obtain a new population. Determine whether to continue updating or output the optimal parameter pair based on whether the preset maximum number of iterations is reached or whether the preset fitness function accuracy is reached. Set the current number of iterations N. When N>Iter max When the maximum number of iterations is reached, or the preset fitness function accuracy is reached, the iteration ends.

[0115] S37: Obtain an optimal parameter pair (C, g), use the optimal parameters to train and model the training set samples, and obtain an output optimal parameter combination;

[0116] S38: Substitute the test set data into the training model after step S38 for testing to obtain the SVM prediction model.

[0117] The present invention performs data preprocessing on the data set of the entire logistics enterprise user, converts the data through the CDF method, and uses the cumulative distribution function to perform dimensionless operations on the data. In the feature engineering stage after preprocessing, the RF and CRITIC methods are used to screen the data and evaluate the user churn coefficient. In the prediction method of user churn, a method combining BAS and SVM is adopted, and the idea of ​​the population and the optimal solution guidance mechanism is added to BAS to accelerate the speed of the longhorn beetle population approaching the optimal individual solution. The penalty factor and kernel function parameters of SVM are better optimized through BAS, which is used to predict user churn and prevent user churn.

[0118] Of course, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting user churn of a logistics enterprise, characterized by: The following steps are involved: S1: Data preprocessing is performed on the original data set of logistics enterprise users. For the original data set, CDF cumulative distribution function is used for data conversion, and all features are graded and classified to make all data dimensionless; S2: Feature engineering stage processing: RF and CRITIC methods are used to extract data features and evaluate user churn coefficients; S3: User churn prediction and processing, using a combination of BAS and SVM, combining the idea of ​​population with the BAS method, adding the idea of ​​population and optimal solution guidance mechanism to BAS, and using BAS to better optimize the penalty factor and kernel function parameters of SVM, making the local and global optimal extreme values ​​more accurate; The step S3 specifically comprises the following steps: S31: Obtain the original data set of enterprise users, set the population size n, and the maximum number of iterations Ite max , set the value range of the optimal parameter set (C (penalty factor), g (kernel function)), divide the data set into two parts, the training set and the test set; S32: Initialize the position of the longhorn beetle group, optimize the penalty factor (C) and kernel parameter (g) of SVM, and the position of each longhorn beetle represents a pair of parameters (C, g); S33: Determine the position coordinates of the left and right whiskers of the longicorn: Where t is the number of loops of the method, ran(k,1) is a random k-dimensional vector, and distance represents the distance between the left and right whiskers of the longicorn. S34: Determine the fitness function. The fitness function selected by this method is as follows: Among them, W i is the current weight factor, X i To obtain real data, To fit the data, calculate the fitness value of each individual beetle and record the optimal value of the current individual and population; S35: Start iteration and update the population position; S36: Calculate the fitness value after the position is updated, update the individual position by comparing the fitness values, and obtain a new population. Determine whether to continue updating or output the optimal parameter pair according to whether the preset maximum number of iterations is reached or the preset fitness function accuracy is reached. The iteration ends when the maximum number of iterations is reached or the preset fitness function accuracy is reached. S37: Obtain an optimal parameter pair (C, g), use the optimal parameters to train and model the training set samples, and obtain an output optimal parameter combination; S38: Substitute the test set data into the training model after step S38 for testing to obtain the SVM prediction model; The step S35 specifically includes the following steps: Optimize the step size parameter in the position update formula and change the original static step size coefficient to a dynamic adjustment coefficient. in, is the maximum step length, Ite is the current number of iterations, Ite max is the maximum number of iterations, Update the position of the longhorn beetle and introduce the MPSO typical optimal solution guidance mechanism to achieve the goal of making the longhorn beetle group approach the optimal individual solution faster. Define the position formula of the longhorn beetle: Among them, Cons is a constant defined between [0,1], Loc best is the position of the optimal individual under the current number of iterations, ran(k,1) is a random k-dimensional vector, and sign(.) is the sign function:

2. The method for predicting user churn of a logistics enterprise according to claim 1, characterized in that: In the step S1, each feature in the original data set is classified and processed, and the features include basic information data, user behavior attribute data and logistics enterprise attribute data. The basic information includes user number, user identity, user scale, and market. The user behavior attribute data includes service status, transaction quantity, and income and expenditure status. The logistics enterprise attribute data includes transaction status and financial subsidies, and each feature is field named and coded.

3. The method for predicting user churn of a logistics enterprise according to claim 2, characterized in that: The CDF cumulative distribution function method is used for data conversion. The specific method is as follows: For the features after hierarchical processing, CDF cumulative distribution function is used to convert the data of each feature. All data in each feature are expressed as x(x1,x2,x3,...,x n ), where n is the total data of each indicator, and the sorting order of each value in the indicator is expressed as (x [1] ,x [2] ,x [3] ,...,x [n] ), the conversion process of indicator data is as follows: For negative indexes, the opposite index number is first used, and then the CDF cumulative distribution function is used to transform the data.

4. The method for predicting user churn of a logistics enterprise according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21: Extract and screen the feature indicators in the data set through RF. RF is composed of multiple decision trees. The generation of each decision tree follows the top-down division principle. The training set is divided from the root node for each decision tree. For each node in the tree, its division attribute is selected based on the principle of minimum purity node, and the training data set is separated at the same time. S22: Use the CRITIC method to assess the risk of user churn.

5. The method for predicting user churn of a logistics enterprise according to claim 4, characterized in that: In step S21, the decision tree generates child nodes during training. When the decision tree meets the branch stopping rule, the growth of the child nodes is stopped, and the "purity" of each feature is quantified. The method for calculating the "purity" of each feature is as follows: S211: Calculate the gain of information entropy in the feature: Where S is the sample set for training, F is the feature contained in the sample, P(F|i) is the probability of each category when the feature appears, and P(i) is the probability of the feature appearing in the sample; S212: Use the features in the sample as random variables and divide the sample S for training: S213: Calculate the "purity" of the features in the sample set and extract the features with high "purity":

6. The method for predicting user churn of a logistics enterprise according to claim 4, characterized in that: The step S22 comprises the following steps: S221: Calculate the conflict quantitative index between the jth characteristic index and other indexes: Among them, the correlation coefficient The covariance and standard deviation of the samples are selected for representation. S222: Calculate the amount of information contained in the jth feature: S223: Calculate the objective weight of the jth feature: Among them, X ij represents the jth secondary indicator in the i-th primary indicator, Represents the average value of the i-th first-level indicator.