A customer information data verification model training method based on decision tree

Abstract

The present invention relates to the field of data processing technology, and in particular to a customer information data verification model training method based on a decision tree. The customer information data verification model training method includes the following processes: data collection to form a data set; collecting customer electricity consumption behavior patterns, historical services, and archival information; screening features that are highly correlated with the accuracy of customer contact information; selecting a decision tree model to train the features after the initial screening; the present invention integrates feature selection, feature classification, and classification prediction, and realizes automatic verification of customer contact information by deeply mining and analyzing customer archival information.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a customer information data verification model training method based on a decision tree. Background Art

[0002] At present, there is a certain proportion of contact mismatch problems in the customer file information of power companies. The main manifestation is that the telephone number registered with the account number is inconsistent with the telephone number of the actual resident customer, which causes the customer service center to contact the wrong user during activities such as collection or business marketing, thereby generating the risk of customer complaints. Therefore, it is necessary to find customers with mismatched numbers as much as possible and verify and improve customer contact information to provide data support. In the absence of automated tools or methods to identify whether the customer's contact information is accurate, business personnel need to first follow defined rules, such as counting the number of customers associated with each contact method, and selecting manual outbound calls and intelligent robot outbound calls according to different frequency divisions, so as to realize full outbound calls to verify whether the contact information is accurate. As shown in the attached Figure 1 As shown, the process is as follows:

[0003] a. Rule predefinition and statistics: Count indicators based on predefined rules, such as counting the number of customers associated with each contact method; b. Select processing methods based on indicator data, such as whether the number of customers associated with the contact method is less than or equal to 5; c. Execute intelligent outbound calls or manual outbound calls: Use the intelligent outbound call system to contact customers or customers with more complicated manual outbound call situations based on different conditions; d. Record and verify: Record the outbound call results. If the outbound call result is wrong, check whether there are other contact methods. If there are other contact methods, repeat the rule judgment of process b; e. Customer verification marking and information maintenance: Mark the customer as verified or unverified based on the correctness of the contact information, and maintain the customer information in the database.

[0004] Manual outbound calls are inefficient and costly. Although intelligent robot outbound calls are more efficient, they may affect the overall customer experience due to the lack of human interaction. Summary of the invention

[0005] The present invention proposes a customer information data verification model training method based on a quantitative model. The specific technical solution is as follows:

[0006] A customer information data verification model training method based on a decision tree includes the following steps:

[0007] S100: Data collection: Through the intelligent outbound calling technology of the customer service center, we collect work orders with incorrect contact information reported by customers during manual collection work, and analyze the characteristics of the error information;

[0008] S200: Based on the customer profile of the target customer, collect the customer's electricity consumption behavior pattern, historical service, and archive information, and form a data set through customer information and outbound call results;

[0009] S300: Perform feature correlation analysis based on the chi-square independence test, conduct preliminary feature screening, calculate the correlation between each indicator and the verification result, and screen features that are highly correlated with the accuracy of customer contact information;

[0010] S400: Based on the preliminary chi-square independence test, a decision tree model is selected to train the model for the features after the initial screening; the specific process includes:

[0011] S410: In the initial feature selection stage, the three indicators of information gain, gain rate and Gini impurity are used to evaluate the contribution of features to the model from multiple angles, so as to effectively divide and select features;

[0012] S420: performing feature selection in a recursive manner, continuously splitting nodes until a specific stop condition is met;

[0013] S430: Systematically traverse multiple hyperparameter combinations through grid search to optimize the generalization ability of the model. The specific hyperparameters include: split strategy splitter, split criterion, maximum depth max_depth, minimum number of split samples min_samples_split, maximum number of features max_features, and minimum number of leaf samples min_samples_leaf;

[0014] S440: The evaluation indicators of the model include: overall accuracy, recall rate, precision rate and outbound call ratio, and they are added to the target score of the grid search to maximize the evaluation indicators. The specific indicator evaluation formula is shown in Table 1 below.

[0015] Table 1 Evaluation index calculation formula

[0016] index Calculation formula Indicator significance Overall accuracy (TP+TN) / (TP+FN+FP+TN) Overall forecast accuracy Recall (R) TN / (TN+FP) The percentage of real error customers that can be found Precision (P) TN / (TN+FN) Measuring model efficiency Outbound call ratio (FN+TN) / (TP+FN+FP+TN) Measure the amount of outbound calls that the model can reduce

[0017] Preferably, the S300 includes the following process:

[0018] S310: Construct the chi-square independence test statement hypothesis,

[0019] Null hypothesis H0: There is no correlation between a customer behavior feature or profile information and the outbound call number and the customer matching category, and they are independent;

[0020] Alternative hypothesis H1: There is a correlation between the feature and the target classification;

[0021] S320: Create a contingency table, construct a matrix to present observations, where one variable defines the rows, another variable defines the columns, and each table cell shows the frequency of observations corresponding to a specific combination of the variable categories;

[0022] S330: Calculate the chi-square statistic, specific formula:

[0023]

[0024] Among them, O is the observed frequency, that is, the frequency in the actual data, and E is the expected frequency. The specific formula is:

[0025] S330: Evaluate the test statistic according to the chi-square distribution to find the p-value, which will indicate the statistical significance of the test; if the p-value is less than the selected significance level, H0 is rejected, indicating that a certain customer behavior characteristic or profile information under test is highly correlated with the target classification.

[0026] Preferably, the S410 includes the following process:

[0027] S411: The specific information gain feature selection method is shown in the following formula:

[0028]

[0029] Among them, p i is the proportion of samples of the i-th category in the data set; H(T) is the entropy of the data set T;

[0030] H(T v ) is a subset T v The entropy of ; Values(A) is the feature set of all features; T v

[0031] It is a subset of the dataset where feature A takes the value v;

[0032] IG(T,A) is the information gain of feature A for data set T;

[0033] S412: The specific gain rate feature selection method is shown in the following formula:

[0034]

[0035]

[0036] Among them, IV(A) is the intrinsic value of feature A, calculated by proportion;

[0037] GR(T,A) is the gain rate of feature A to the data set;

[0038] S413: The specific Gini impurity feature selection method is shown in the following formula:

[0039]

[0040] Among them, Gini(T) is the Gini impurity of data set T; Gini(T v ) is the data set T v Gini impurity of ;

[0041] Gini_gain(T,A) is the reduction in the Gini impurity of dataset T caused by feature A.

[0042] The advantages of the present invention over the prior art are: the present invention constructs a model based on a decision tree algorithm by deeply analyzing multiple dimensional information in customer files, aiming to identify and determine whether there are incorrect contact information in customer files. The present invention integrates feature selection, feature classification, and classification prediction, and realizes automatic verification of customer contact information by deeply mining and analyzing customer file information. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The present invention is a flow chart of a data verification method in the prior art.

[0044] Figure 2 It is a flow chart of the data verification model training method of the present invention.

[0045] Figure 3 It is a schematic diagram of the process flow of the present invention.

[0046] Figure 4 It is a specific decision tree summary of the model of the present invention. DETAILED DESCRIPTION

[0047] A customer information data verification model training method based on a decision tree includes the following steps:

[0048] S100: Data collection: Through the intelligent outbound calling technology of the customer service center, we collected the work orders of incorrect contact information reported by customers in the early stage of manual collection of fees for low-voltage residential customers. We extracted the customer account number, outbound phone number and other information from the work orders, and analyzed the characteristics of the error information;

[0049] S200: Based on the customer portrait of the target customer, collect the customer's electricity consumption behavior pattern, historical service, and archive information, and form a data set through customer information and outbound call results, where the characteristics organized in combination with business input are shown in Table 2 below;

[0050] Table 2 Feature names and feature dimensions

[0051] Feature Name Feature Dimension Is the outbound call number from the online State Grid? Outbound call number Number of mobile phones in the account number file information Account Number Number of fixed telephone numbers in the account number file information Account Number Outbound call number appears in the archive Outbound call number The outbound call number appears in the online national network registration number Outbound call number The outbound call number appears in the subscription account Outbound call number The outbound call number appears in the debit account Outbound call number The outbound number comes from the customer file Outbound call number The outbound call number comes from the customer's online State Grid Outbound call number The outbound call number comes from the customer's subscription number Outbound call number The outbound call number comes from the customer's debit account Outbound call number Number of business operations of account number information maintenance in the past 90 days User Conduct Number of account name change and transfer transactions in the past 90 days User Conduct Total number of account number services received in the past 90 days User Conduct Total number of contact details in the account number file information Account Number Account Number Year of Establishment Account Number The ID card under the account number has not been verified. Account Number The ID card under the account number has been verified. Account Number When the account number is bound to the online State Grid account, the telephone number of the account number that matches the online State Grid account Account Number When the account number has a subscription record, the number of numbers that match the account number and subscription number Account Number When the account number is used for deduction, the number of the account number that matches the deduction account number Account Number Total number of accounts associated with outbound calls telephone number Is the account number bound to a non-file number? Account Number Whether the account number can be used for multiple payment channels Account Number Number of mobile phones in the account number file information Account Number Number of times the outbound number has called 95598 in the past 180 days telephone number How many online State Grid accounts have the account number been bound to? Account Number How many subscription accounts are bound to the account? Account Number Number of unregistered telephone numbers of the State Grid on the Internet Account Number Number of online State Grid telephone numbers with real names Account Number Is the account number deducted? Account Number

[0052] S300: Based on the chi-square independence test, we conducted feature correlation analysis, carried out preliminary feature screening, calculated the correlation between each indicator and the verification result, and screened features that were highly correlated with the accuracy of customer contact information, such as the legitimacy of customer names, the diversity of contact information, etc. However, before conducting the chi-square independence test analysis, since the chi-square test is mainly used to evaluate the independence between categorical variables, it is necessary to ensure that all features involved in the analysis are discrete features. Therefore, we merged the above features, performed a chi-square independence test on the processed discrete features, and calculated the corresponding chi-square value and P value; the following process is included:

[0053] S310: Construct the chi-square independence test statement hypothesis,

[0054] Null hypothesis H0: There is no correlation between a customer behavior characteristic or profile information and the classification of outbound call numbers and customer matching, and they are independent;

[0055] Alternative hypothesis H1: There is a correlation between the feature and the target classification;

[0056] S320: Create a contingency table, construct a matrix to present observations, where one variable defines the rows, another variable defines the columns, and each table cell shows the frequency of observations corresponding to a specific combination of the variable categories;

[0057] S330: Calculate the chi-square statistic, specific formula:

[0058]

[0059] Among them, O is the observed frequency, that is, the frequency in the actual data, and E is the expected frequency. The specific formula is:

[0060] S330: Evaluate the test statistic according to the chi-square distribution to find the p-value, which will indicate the statistical significance of the test; if the p-value is less than the selected significance level, p is usually 0.05, then H0 is rejected, indicating that a certain customer behavior characteristic or profile information under test is highly correlated with the target classification;

[0061] The specific chi-square independence test results are shown in Table 3 below:

[0062] Table 3 Chi-square independence test results

[0063]

[0064] Note: Due to the length of P values, the P values ​​shown in the table are only retained to three decimal places.

[0065] S400: According to the preliminary chi-square independence test, the features with a P value less than or equal to 0.05 are selected as the subsequent training features of the decision tree model. At the same time, in order to accurately define the rules for subsequent decision tree classification, the processed chi-square features are restored and the original features are trained. The specific process includes:

[0066] S410: Feature selection. In order to comprehensively improve the accuracy of the model, the present invention uses three indicators, namely information gain, gain rate and Gini impurity, in the initial feature selection stage to evaluate the contribution of features to the model from multiple angles, thereby effectively dividing and selecting features. The process includes the following:

[0067] S411: The specific information gain feature selection method is shown in the following formula:

[0068]

[0069] Among them, p i is the proportion of samples of the i-th category in the data set; H(T) is the entropy of the data set T;

[0070] H(T v ) is a subset T v The entropy of ; Values(A) is the feature set of all features; T v

[0071] It is a subset of the dataset where feature A takes the value v;

[0072] IG(T,A) is the information gain of feature A for data set T;

[0073] S412: The specific gain rate feature selection method is shown in the following formula:

[0074]

[0075] Among them, IV(A) is the intrinsic value of feature A, calculated by proportion;

[0076] GR(T,A) is the gain rate of feature A to the data set;

[0077] S413: The specific Gini impurity feature selection method is shown in the following formula:

[0078]

[0079] Among them, Gini(T) is the Gini impurity of data set T; Gini(T v ) is the data set T v Gini impurity of ;

[0080] Gini_gain(T,A) is the reduction of feature A on the Gini impurity of data set T;

[0081] S420: performing feature selection in a recursive manner, continuously splitting nodes until a specific stop condition is met;

[0082] S430: Systematically traverse multiple hyperparameter combinations through grid search to optimize the generalization ability of the model. The specific hyperparameters include: splitting strategy splitter, splitting criterion criterion, maximum depth max_depth, minimum number of split samples min_samples_split, maximum number of features max_features, minimum number of leaf samples min_samples_leaf. These hyperparameters involve pre-pruning or post-pruning strategies, which can effectively prevent model overfitting.

[0083] S440: The evaluation indicators of the model include: overall accuracy, recall rate, precision rate and outbound call ratio, and they are added to the target score of the grid search to maximize the evaluation indicators. The specific indicator evaluation formula is shown in Table 1.

[0084] Test set verification: The present invention divides the training set and the test set into a ratio of 7:3, verifies the classification effect of the model on the test set, and calculates the evaluation index.

[0085] Effect comparison: The decision trees produced by different evaluation indicators were compared with the full outbound call model and the random parameter model, and it was found that the performance of the set analysis model was significantly better than the general method. Among them, there are 5 specifically designed analysis models, which optimize the five scenarios of "combined value of total accuracy and precision", "recall rate", "outbound call share", "combined value of recall rate and precision", and "combined value of recall rate, precision rate and outbound call share", with the goal of improving accuracy, precision, recall rate and reducing outbound call share as much as possible. The specific results are as follows:

[0086] Table 4 Comparison of analytical model performance

[0087] Optimization goals Prediction accuracy Prediction accuracy Prediction Recall Outbound call ratio Full outbound call (comparison) 100.00％ 13.13％ 100.00％ 100.00％ Random parameter model (control) 67.54％ 18.90％ 44.72％ 31.08％ Overall accuracy and precision 67.37％ 21.62％ 56.52％ 34.34％ Recall 48.94％ 18.45％ 84.47％ 60.11％ Outbound call ratio 77.73％ 26.47％ 39.13％ 19.41％ Recall and Precision 61.34％ 19.85％ 63.98％ 42.33％ Recall, precision, and call share 50.24％ 18.86％ 84.47％ 58.81％

[0088] Without considering the busy period of the customer service center, the model that optimizes the "combined value of the three indicators of recall rate, precision rate and outbound call ratio" can integrate multiple factors. Compared with the current full outbound call mode, the model can identify 84.47% of wrong customers and reduce the outbound call workload to the current level of 58.81%. Therefore, in daily customer service outbound call work, the present invention will also promote this analysis model. The specific decision tree outline is as follows Figure 4 shown.

[0089] This analysis model is deployed locally, and the specific deployment parameters are: class weight class_weight = balanced to reduce the impact of data imbalance; splitting criterion criterion = gini; maximum depth max_depth = 5; maximum number of features max_features = log2; maximum number of leaf nodes max_leaf_nodes = none; minimum reduction in node impurity min_impurity_decrease = 0.0; minimum number of leaf samples min_samples_leaf = 14; minimum number of samples for min_samples_split = 2; minimum weight fraction of leaf nodes min_weight_fraction_leaf = 0.0; splitting strategy splitter = random.

[0090] In order to study the effectiveness of the model in the future, a pilot project was conducted in Phoenix Power Supply Company in Xingtai, Hebei Province. Based on the 1,375 customers provided by the local companies, data on their characteristics in various dimensions was collected and input into the decision tree model as a data set. Based on the model's predictive reasoning, an assessment was made on whether outbound calls were needed. Outbound calls were made to 873 customers who were recommended to make outbound calls, and a total of 107 customers with incorrect contact information were found. Subsequently, in order to verify the recall rate of the model, the customer service center also made outbound calls to 502 customers who were recommended by the model not to make outbound calls, and a total of 12 customers with incorrect contact information were found. According to this calculation, when the model is used in actual work, it can reduce the workload of outbound calls by 36.51% and identify 89.92% of customers with incorrect contact information, achieving the expected effect.

Claims

1. A customer information data verification model training method based on a decision tree, characterized in that: The process includes the following: S100: Data collection: Through the intelligent outbound calling technology of the customer service center, we collect work orders with incorrect contact information reported by customers during manual collection work, and analyze the characteristics of the error information; S200: Based on the customer profile of the target customer, collect the customer's electricity consumption behavior pattern, historical service, archive information, customer archive information and outbound call results to form a data set; S300: Perform feature correlation analysis based on the chi-square independence test, conduct preliminary feature screening, calculate the correlation between each indicator and the verification result, and screen features that are highly correlated with the accuracy of customer contact information; S400: Based on the preliminary chi-square independence test, a decision tree model is selected to train the model for the features after the initial screening; The specific process includes: S410: In the initial feature selection stage, the three indicators of information gain, gain rate and Gini impurity are used to evaluate the contribution of features to the model from multiple angles, so as to effectively divide and select features; S420: performing feature selection in a recursive manner, continuously splitting nodes until a specific stop condition is met; S430: Systematically traverse multiple hyperparameter combinations through grid search to optimize the generalization ability of the model. The specific hyperparameters include: splitting strategy (splitter), splitting criterion (criterion), maximum depth (max_depth), minimum number of split samples (min_samples_split), maximum number of features (max_features), and minimum number of leaf samples (min_samples_leaf); S440: The evaluation indicators of the model include: overall accuracy, recall rate, precision rate and outbound call ratio, and they are added to the target score of the grid search to maximize the evaluation indicators. The specific indicator evaluation formula is shown in Table 1 below. Table 1 Evaluation index calculation formula 2. According to claim 1, a customer information data verification model training method based on a decision tree is characterized in that: The S300 includes the following process: S310: Construct the chi-square independence test statement hypothesis, Null hypothesis H0: There is no correlation between a customer behavior characteristic or profile information and the classification of outbound call numbers and customer matching, and they are independent; Alternative hypothesis H1: There is a correlation between the feature and the target classification; S320: Create a contingency table, construct a matrix to present observations, where one variable defines the rows, another variable defines the columns, and each table cell shows the frequency of observations corresponding to a specific combination of the variable categories; S330: Calculate the chi-square statistic, specific formula: Among them, O is the observed frequency, that is, the frequency in the actual data, and E is the expected frequency. The specific formula is: S330: Evaluate the test statistic according to the chi-square distribution to find the p-value, which will indicate the statistical significance of the test; if the p-value is less than the selected significance level, H0 is rejected, indicating that a certain customer behavior characteristic or profile information under test is highly correlated with the target classification.

3. According to claim 1, a customer information data verification model training method based on decision tree is characterized in that: The S410 includes the following process: S411: The specific information gain feature selection method is shown in the following formula: Among them, p i is the proportion of samples of the i-th category in the data set; H(T) is the entropy of the data set T; H(T v ) is a subset T v The entropy of ; Values(A) is the feature set of all features; T v It is a subset of the data set where feature A takes the value v; IG(T,A) is the information gain of feature A for data set T; S412: The specific gain rate feature selection method is shown in the following formula: Among them, IV(A) is the intrinsic value of feature A, calculated by proportion; GR(T,A) is the gain rate of feature A to the data set; S413: The specific Gini impurity feature selection method is shown in the following formula: Among them, Gini(T) is the Gini impurity of data set T; Gini(T v ) is the data set T v Gini impurity of ; Gini_gain(T,A) is the reduction in the Gini impurity of dataset T caused by feature A.

Images