CN110019166A - Method for screening attribute data and customer loss early warning method - Google Patents

Abstract

A method for screening attribute data and a customer loss early warning method belong to the field of data processing, and the method for screening the attribute data comprises the following steps: screening attribute data by using an information gain method; screening attribute data by using a point biserial correlation coefficient method; according to the information gain method and the point biserial coefficient method, a plurality of attribute data are respectively obtained, and the two attribute data are intersected to obtain the screened attribute data, so that the effect is as follows: the method combines the information gain method and the point biserial correlation coefficient method, so that the screening accuracy is better.

Description

The method of screening attribute data and the method of early warning of customer churn

technical field

The invention belongs to the field of data processing, and relates to a method for screening attribute data and a method for early warning of customer loss

Background technique

At present, with the continuous increase in the number of car dealers and repair stations, the competition in the automotive after-sales service market has become more and more fierce. In addition, the sensitivity of customers to service prices and the continuous improvement of customer satisfaction with in-store services have led to the loss of customers in 4S stores. The rate is rising year by year, and the direct consequence of the loss of a large number of customers is the economic loss. Thinking from a deeper level, this will also indirectly affect the reputation of the 4S store, thus falling into a vicious circle of declining revenue and reputation. So thinking from the perspective of 4S stores, how to effectively identify customers with a high probability of loss, and how to successfully achieve customer retention are the top priorities for good customer relationship management. In addition, with the continuous maturity of data mining technology and the continuous accumulation of 4S store operation data, it provides a good foundation for exploring the possibility of customer churn from the perspective of data. In this paper, a method is invented. On the basis of in-depth use of 4S store operation data, a customer churn warning model based on decision tree algorithm is proposed. Through this model, 4S stores can grasp the list of customers with high loss probability in the future. So as to provide good conditions for customer relationship maintenance.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention proposes the following scheme: a method for screening attribute data, comprising: screening attribute data using an information gain method; screening attribute data using a point double column correlation coefficient method; according to the information gain method and the point double column coefficient The method obtains several attribute data respectively, and takes the intersection of the two to obtain the filtered attribute data.

The present invention also provides a method for early warning of customer loss, which is characterized by comprising: S1. collecting basic attribute data of customers, purchasing vehicle data and after-sale behavior data of entering 4S stores; S2. determining target variables and independent variables; S3. Screening independent variables; S4. Constructing a decision tree model; S5. Using the decision tree model to actually predict, and issuing a loss alarm when necessary; wherein, screening the independent variables, screening with the method for screening attribute data described in any one of the present invention .

Beneficial effect: Combining the two independent variable screening methods of the information gain method and the point double column correlation coefficient method provides a new idea for the variable screening method of the classification model to improve the screening accuracy. The attribute has the characteristics of customer relationship management in the automotive industry, so that the constructed decision tree model has more industry applicability, and provides a feasible customer churn early warning scheme for customer relationship management in the automotive industry.

Description of drawings

Figure 1 shows the flow chart of customer churn early warning.

Detailed ways

The present invention mainly realizes through the following technical solutions:

1. Collect the basic attribute data of each customer, purchase vehicle data and after-sale 4S store behavior data, and build a database

1) The basic attribute data of customers mainly include ID number, name, gender, age, province, city, contact information, education level, hobbies, industry and other information;

2) Purchased vehicle data mainly includes chassis number, dealer, model, price, etc.;

3) The after-sales behavior data of entering 4S stores mainly include the type of entry (such as repair, maintenance, insurance, claim, etc.), entry time, entry consumption amount, entry mileage, labor cost, spare parts fee, settlement date, and maintenance items. Wait.

2. Data cleaning

1) Missing value processing: For example, missing gender, age, province, and city can be translated and supplemented by the corresponding digits of the ID number; missing models can be translated and supplemented based on the chassis number; missing prices can be based on the car model. The average value of the selling price is supplemented; the missing consumption amount can be supplemented after calculation according to the maintenance items and the unit price of the item;

2) Identification and processing of noise data: Due to the reasons of data collection and input, there will be some noise data in the customer's after-sales data of entering the store. First, the part of the data needs to be identified. In the present invention, the DBSCAN algorithm (a density-based clustering algorithm) is mainly used to identify the noisy data; secondly, for the identified outliers, the binning method is used to "smooth" the data.

3. Determine the target variable

The target variable (churn or non-churn) is determined according to whether the customer enters the 4S store within the specified time. If the customer enters the 4S store within the specified time, it is a non-churn customer, otherwise it is a lost customer. The commonly used time window is 3 months, 6 month or year, and one year is used as a time window in the present invention.

4. Calculate the independent variable

According to the data after data cleaning obtained in step 2, calculate several basic attributes related to customer churn, mainly including age, price, car age, mileage of the last 4S store entry, number of times of entering 4S stores, number of trips, Cumulative accident costs, maintenance times, average annual maintenance times, average maintenance costs per time, overdue maintenance times, cumulative maintenance costs, maintenance times, average annual maintenance times, average maintenance costs per time, and cumulative maintenance costs.

5. Filter independent variables

The importance of the independent variables in step 4 is evaluated by combining the information gain method and the point double-column correlation coefficient method, so as to screen out several attributes with high importance.

Among them, the main process of the information gain method is as follows:

1) Calculate the expected information Info(D) that correctly classifies the observations in D

where D represents all observation data sets, p _i is the non-zero probability that any observation in D belongs to class C _i , and is estimated by |C _{i, D} |/|D|;

2) Calculate the amount of information required to classify observations in D according to attribute A

Among them, attribute A has v different values {a ₁ ,a ₂ ,...,α _v } according to data set D, and D can be divided into v sub-region subsets {D ₁ ,D ₂ ,...,D by attribute A _v }, where D _j contains observations in D whose A value is α _j . 3) Calculate the information gain of attribute A

Gain(A)=Info(D)-Info _A (D)

4) Set the threshold to remove the basic attribute with small information gain

The main process of the point double column correlation coefficient method is as follows:

1) Calculate the variable ratio Y _p that takes a certain value in the target variable Y and the variable ratio Y _q that takes another value;

2) Calculate the mean value of the part corresponding to Y _p in the independent variable X

3) Calculate the mean value of the part corresponding to Y _q in the independent variable X

4) Calculate the standard deviation S _x of the independent variable X;

5) According to the formula Calculate the correlation coefficient between the independent variable X and the target variable Y.

According to the information gain method and the point double column coefficient method, several independent variables with high importance can be obtained respectively, and the comprehensive results of the two methods can be obtained by taking the intersection of the two methods.

After the screening of the above methods, the independent variables with high importance are obtained including the mileage of the last 4S store, the number of accidents, the price, the average number of annual maintenance, the average maintenance cost, and the number of overdue maintenance.

6. Build a decision tree model

Use the independent variables screened in step 5 to build a decision tree model, calculate the information gain of each attribute, select the attribute with the largest information gain as the root node, and establish a branch for each value of the attribute. branch, calculate the information gain for all the remaining attributes, also select the attribute with the largest information gain as the new split node and establish the corresponding branch, recurse the above process until there is no attribute, define the node as a leaf node, and assign the Nodes are marked with the most numerous categories among all samples. In addition, leaf nodes without any samples need to be pruned.

7. Verify the accuracy of the model

Input the customer data marked with churn and non-churn labels into the constructed decision tree model for analysis, compare the difference between the predicted results and the actual results, so as to determine the accuracy of the model and revise the model.

8. Actual forecast, release early warning of churn

According to the revised churn prediction model, the current non-churn customers are predicted, focusing on customers with a high churn probability, and issuing churn warnings.

Claims (9)

1. a kind of method for screening attribute data characterized by comprising

Use information gain method screens attribute data；

Attribute data is screened using point biserial correlation coefficient method；

Several attribute datas are respectively obtained according to information gain method and point biserial coefficient method, and after the two takes intersection to obtain screening Attribute data.

2. the method for screening attribute data as described in claim 1, which is characterized in that information gain method screens attribute data The step of method, comprising:

1) is calculated in D and is observed the expectation information Info (D) correctly to classify

Wherein, D indicates all observation data sets, p_iIt is that arbitrarily observation belongs to class C in D_iNonzero probability, be used in combination | C_{I, D}|/| D | estimate Meter, i indicate the value serial number of observation, and m indicates observation total amount.

2) it calculates and required information content of classifying is carried out to observation in D according to attribute A

Wherein, attribute A has v different value { α according to data set D₁,α₂,…,α_v, D is divided into v subregion with attribute A Subset { D₁,D₂,…,D_v, D_jComprising the observation in D, their A value is α_j。

3) information gain of computation attribute A

Gain (A)=Info (D)-Info_A(D)；

4) given threshold removes the basic attribute data of part according to gain, removes the basic attribute data removed in D, remaining Basic attribute data is to be screened out attribute data.

3. the method for screening attribute data as described in claim 1, which is characterized in that point biserial correlation coefficient method screens attribute The method of data, comprising:

1) the variable scale Y that certain value is taken in target variable Y is calculated_pAnd take the variable scale Y of an other value_q；

2) calculate independent variable X in Y_pThe average value of corresponding part

3) calculate independent variable X in Y_qThe average value of corresponding part

4) the standard deviation S of independent variable X is calculated_x；

5) according to formulaTo calculate the related coefficient of independent variable X and target variable Y.

4. the method for screening attribute data as described in claim 1, which is characterized in that use the attribute data after screening, structure The method for building decision-tree model is: selecting the maximum attribute of information gain as root node, and builds for each value of the attribute A branch is found, for each branch, the maximum attribute of information gain in other remaining all properties is selected to divide as new Node is split, and establishes corresponding branch, recurrence above procedure, until no any attribute, defining the node is leaf node, and will The vertex ticks is the largest number of classifications in all samples.

5. a kind of customer defection early warning method characterized by comprising

S1. it acquires the basic attribute data of client, buy vehicle data and after sale into the shop 4S behavioral data；

S2. target variable, independent variable are determined；

S3. independent variable is screened；

S4. decision-tree model is constructed；

S5. decision-tree model actual prediction is used, and publication is lost alarm if necessary；

Wherein, independent variable is screened, is screened in the method for screening attribute data of any of claims 1-3.

6. customer defection early warning method as claimed in claim 5, which is characterized in that the method for the building decision-tree model, Use the method for the building decision-tree model in the method for screening attribute data as claimed in claim 4.

7. customer defection early warning method as claimed in claim 5, which is characterized in that step S1 includes that step S1_2. data are clear It washes, the data cleansing includes one or more of following methods:

1) missing values are handled；

2) identification and processing of noise data.

8. customer defection early warning method as claimed in claim 5, which is characterized in that further comprise the steps of: the standard of S6. verifying model True property: mark is lost, the decision-tree model that the input of the customer data of non-streaming lose-submission label is built is analyzed, comparison prediction knot The difference of fruit and actual result, to determine the accuracy of model and be modified to model.

9. the customer defection early warning method as described in claim 5 or 8, which is characterized in that it further include step S6. actual prediction, Publication is lost early warning: being predicted according to decision-tree model or revised decision-tree model current non-attrition customer, needle To the higher client of probability is lost, publication is lost early warning.