KR20190134934A - Predictive device for customer churn using Deep Learning and Boosted Decision Trees and method of predicting customer churn using it - Google Patents

Abstract

The present invention relates to a device for predicting customer churn, configured to predict customer churn from a service provided by a company. According to an embodiment of the present invention, the device for predicting customer churn comprises: an input unit to which customer data is inputted; a pre-processing unit configured to pre-process the customer data inputted to the input unit and classify the customer data into time-based pre-processed data and customer-based pre-processed data; a customer churn prediction unit configured to extract a behavior pattern using the classified time-based pre-processed data and a statistical index of a customer using the classified customer-based pre-processed data, and predict customer churn by using the extracted behavior pattern and statistical index; and an output unit configured to output prediction results of the customer churn prediction unit. According to the present invention, it is possible to predict customer churn with higher accuracy.

Description

Prediction device for customer churn using Deep Learning and Boosted Decision Trees and method of predicting customer churn using it}

The present invention relates to a customer departure prediction device for predicting whether a customer will leave the service provided by the company and a customer departure prediction method using the same.

Many companies recognize the importance of customer-centric business strategies in order to stay ahead of the competition and generate stable returns. Because companies rely on revenue from their customers, customer relationship management makes decisions based on data that reflects their customers' actual behavior. For example, data analysis can assess the potential value of customers and predict the likelihood of customer churn.

Customer churn, defined as the tendency of customers to stop trading, is one of the major challenges facing companies around the world. Reducing sales from customer churn not only creates opportunity costs, but also increases the need for new customers. However, the cost of acquiring new customers is estimated to range from $ 300 to $ 600, which is five to six times higher than the cost of retaining existing customers.

Therefore, predicting and preventing existing customers who are more likely to drop out is more cost effective than attracting new customers. In order to effectively manage customer churn, it is important to establish accurate customer churn prediction models.

A Study on the Customer Departure Prediction Model of Department Stores through Data Mining Techniques

An object of the present invention is to provide a customer departure prediction apparatus that can determine whether the customer is separated with improved accuracy and a customer departure prediction method using the same.

Customer churn prediction apparatus according to an embodiment of the present invention,

An input unit to which customer data is input; A pre-processing unit for pre-processing the customer data input to the input unit and classifying the time-based pre-processing data and the customer-based pre-processing data; The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting customer churn prediction unit; And an output unit for outputting a prediction result of the customer deviation predicting unit.

In the customer departure prediction apparatus according to an embodiment of the present invention, the pre-processing unit, the time-based pre-processing module for pre-processing the customer data input to the input unit on a time basis to classify the time-based pre-processing data, the customer input to the input unit It includes a customer-based preprocessing module that pre-processes the data to customer standards and classifies the data into customer-based preprocessing data.

In the customer departure predicting apparatus according to an embodiment of the present invention, the time-based preprocessing data is in a time table format in which customer behaviors are arranged in chronological order, and the customer-based preprocessing data includes a plurality of customer attributes regardless of time. It can be an formatted attribute table format.

In the customer deviation prediction apparatus according to an embodiment of the present invention, the customer deviation prediction unit extracts a behavior pattern by using the classified time-based pre-processing data to predict whether or not the customer departure, and the classification Statistical index extraction model using the pre-processed customer-based preprocessed data to predict the departure of customers and predicting more than half by applying the voting technique to the prediction results of the behavior pattern extraction model and statistical index extraction model It includes an ensemble prediction model that produces a result.

In the customer departure prediction apparatus according to an embodiment of the present invention, the behavior pattern extraction model is a prediction model based on RNN (Recurrent Neural Network, cyclic neural network), a convolutional neural network (convolutional neural network) based prediction model, the The statistical index extraction model may be a prediction model based on a random forest (RF) or a prediction model based on extreme gradient boosting (XGBoost).

In the customer churn prediction apparatus according to an embodiment of the present invention, the customer churn prediction unit may be composed of at least three prediction models of the RNN, CNN, RF, XGBoost-based prediction model.

In the customer departure prediction apparatus according to an embodiment of the present invention, the ensemble prediction model may include the RNN based prediction model, the RF based prediction model, and the XGBoost based prediction model.

Customer churn prediction method according to an embodiment of the present invention,

A first step performed by the computing means and inputting customer data; A second step of pre-processing the input customer data and classifying it into time-based preprocessing data and customer-based preprocessing data; The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting a third step; And a fourth step of outputting a prediction result of the customer deviation predicting unit.

In the customer departure prediction method according to an embodiment of the present invention, the second step, pre-processing the input customer data on a time basis to classify the time-based pre-processing data, by pre-processing the input customer data on a customer basis Classify as customer standard preprocessing data.

In the customer departure prediction method according to an embodiment of the present invention, the time-based preprocessing data is in a time table format in which customer behaviors are arranged in chronological order, and the customer-based preprocessing data includes a plurality of customer attributes irrespective of time. It can be an formatted attribute table format.

In the method of predicting customer departure according to an embodiment of the present invention, the third step includes: extracting a behavior pattern by using the classified time-based preprocessing data to predict whether the customer is leaving, and the classified customer standard. The method includes extracting statistical indicators using preprocessing data to predict whether the customer deviates, and calculating a prediction result of more than half by applying a voting technique to the prediction result.

In the customer churn prediction method according to an embodiment of the present invention, the behavior pattern may be extracted using an RNN based prediction model, a CNN based prediction model, and the statistical indicator using an RF based prediction model or an XGBoost based prediction model. Can be.

In the customer departure prediction method according to an embodiment of the present invention, the third step may use at least three prediction models of the RNN, CNN, RF, and XGBoost based prediction models.

In the customer churn prediction method according to an embodiment of the present invention, the process of calculating a majority or more prediction results by applying a voting technique to the prediction results, the RNN-based prediction model and the RF-based prediction model, and It is preferable to use an XGBoost based prediction model.

A computer program stored in a recording medium according to an embodiment of the present invention,

At the computer, a first step of entering customer data; A second step of pre-processing the input customer data and classifying it into time-based preprocessing data and customer-based preprocessing data; The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting a third step; And outputting a prediction result of the customer deviation predicting unit.

Other specific details of embodiments according to various aspects of the present invention are included in the following detailed description.

According to the customer churn prediction apparatus according to an embodiment of the present invention, it is possible to determine whether the customer churn with improved accuracy.

1 is a block diagram illustrating a customer churn prediction apparatus according to an embodiment of the present invention.
2 is an example of time-based preprocessing data in a time table format in which customer behaviors are arranged in chronological order.
3 is an example of customer-based preprocessing data in an attribute table format in which a plurality of customer attributes are arranged.
4 is a diagram illustrating an RNN model for extracting a behavior pattern of a customer.
5 is a diagram illustrating a CNN model for extracting a behavior pattern of a customer.
6 is a data set used for improved accuracy measurement of a customer churn prediction apparatus according to an embodiment of the present invention.
FIG. 7 is a table illustrating the accuracy of a result of predicting the data set of FIG. 6 using various prediction models.
8 is a table illustrating the accuracy of each prediction model.
9 is a graph illustrating the importance of attributes.

As the invention allows for various changes and numerous embodiments, particular embodiments will be illustrated and described in detail in the detailed description. However, this is not intended to limit the present invention to specific embodiments, it should be understood to include all transformations, equivalents, and substitutes included in the spirit and scope of the present invention.

The terminology used herein is for the purpose of describing particular example embodiments only and is not intended to be limiting of the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In the present invention, the terms 'comprise' or 'have' are intended to indicate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, and one or more other features. It is to be understood that the present invention does not exclude the possibility of the presence or the addition of numbers, steps, operations, components, components, or a combination thereof.

The customer churn prediction apparatus of the present invention uses a deep learning-based prediction model that predicts customer churn using a customer behavior pattern and a boosted decision tree-based prediction model that utilizes statistical characteristics of the customer. The customer churn prediction apparatus of the present invention discloses a data preprocessing technique for effective customer churn prediction of each model, and an ensemble model that utilizes both customer behavior patterns and statistical characteristics of customers. Hereinafter, a customer departure prediction apparatus and a customer departure prediction method according to an embodiment of the present invention will be described with reference to the drawings.

1 is a block diagram illustrating a customer departure prediction apparatus according to an embodiment of the present invention. As illustrated in FIG. 1, the apparatus for estimating customer deviation according to an embodiment of the present invention includes an input unit 100, a preprocessor 200, a customer deviation predictor 300, and an output unit 400.

The input unit 100 receives customer data. Customer data is a variety of data used for estimating customer churn, for example, in the game industry may be data indicating the log-in data of a particular game user and the activity history of the game, and in the financial industry It may be a history of use of various cards in use.

The preprocessing unit 200 preprocesses the customer data input to the input unit 100 and classifies the data into time-based preprocessing data and customer-based preprocessing data. The preprocessor 200 includes a time-based preprocessing module 210 and a customer-based preprocessing module 220, and the time-based preprocessing module 210 preprocesses the customer data input to the input unit 100 on a time basis. The pre-processing data is classified, and the customer standard preprocessing module 220 pre-processes the customer data input to the input unit 100 based on the customer standard and classifies it as the customer standard preprocessing data.

The time-based preprocessing data classified by the time-based preprocessing module 210 is data reflecting a change in customer behavior over time.

As shown in FIG. 2, the time-based preprocessing data may be in a time table format in which customer behaviors are arranged in chronological order. In Tables 2a and 2b, “Customer ID” refers to a customer ID and “Time” refers to a time when a customer performs a specific action. “Attribute_standard” means standard attribute. The reference attribute is a representative attribute that is a standard of customer behavior. For example, when a customer plays a game, the reference attribute may be one of several characters provided in the game. In the table (a), character 1 is the reference attribute, and in table (b) the character 2 is the reference attribute. Also, for example, in the case of a card used by a customer, “Attribute_standard” may be two A cards 1 and B cards 2 of different types in the tables (a) and (b) of FIG. 2.

“Attribute1” and “Attribute2” refer to behavior attributes that represent customer behavior. For example, in the case of a game, the second line of the table (a) shows the amount of game money (Attribute1) and items (Attribute2) acquired by the customer (4175D6K2) as a character on March 4, 2017. Means.

As such, the behavior attribute that can be expressed as “Attribute + number” can be defined in various ways. For example, the amount of experience that a character of the user (the customer) gets every second, the amount of money that the character of the user (the customer) gets every second, the path that the character of the user (the customer) moves every second, every second It may be an action (attack, skill play, etc.) that the character of the user (customer) takes.

Since the time-based preprocessing data expresses the behavior change of the customer over time as it is, it is effective in extracting the customer's behavior pattern.

As described above, the behavior pattern of the customer may be expressed by subdividing based on specific attributes. For example, the behavior pattern of a customer in the financial industry can be separated into the payment pattern of each card in use. The behavior pattern of a customer in the game industry can be separated into the behavior pattern of each character owned by the customer. As such, by dividing the behavioral data of the customer by using the reference attribute, more sophisticated customer behavioral patterns can be extracted.

The time base preprocessing module 210 classifies the customer's behavior data based on the base attribute and classifies the time base preprocessing data over time. The time-based preprocessing data classified by the time-based preprocessing module 210 is used to construct the behavior pattern extraction model 310.

The customer-based preprocessing data classified by the customer-based preprocessing module 220 is data reflecting statistical characteristics of the customer over time.

The customer-based preprocessing data may be in an attribute table format in which a plurality of customer attributes are arranged regardless of time, as shown in FIG. 3. The customer-based preprocessing data is the data representing the characteristic attributes of each customer, and includes the average value, the maximum value, and the minimum value of the customer's attributes. For example, when the customer plays a specific game in a day as a target attribute, the customer-based preprocessing data may be an average value, a maximum value, a minimum value, or the like of the time the customer (game user) played in a day.

In the table of FIG. 3, “Customer ID” means a customer ID, “Attribute1_mean” means an average value of Attribute 1, “Attribute2_mode” means a maximum / minimum value of Attribute 2, and “Attribute3_slope” means a slope of Attribute 3, and the like. In the table of FIG. 3, "Attribute1_mean", "Attribute2_mode", and "Attribute3_slope" are arranged for different attributes, respectively, but when it is about the same attribute of the same customer ID, "Attribute2_mode" is "Attribute1_mode", "Attribute3_slope" May be arranged as “Attribute1_slope”.

The customer-based preprocessing data classified by the customer-based preprocessing module 220 is used to build a statistical index extraction model 320.

The customer departure prediction unit 300 predicts whether the customer has left by using the time-based preprocessing data and the customer-based preprocessing data complementarily. The customer deviation predictor 300 includes a behavior pattern extraction model 310, a statistical index extraction model 320, and an ensemble prediction model 330.

In the behavior pattern extraction model 310, the behavior pattern is extracted using the time-based preprocessing data to predict the departure of the customer, and in the statistical index extraction model 320, the statistical index is extracted using the customer-based preprocessing data. Predict whether or not a deviation will occur. The ensemble prediction model 330 applies a voting technique to the prediction results of the behavior pattern extraction model 310 and the statistical index extraction model 320 to calculate more than half of the prediction results.

The behavior pattern extraction model 310 extracts the behavior pattern of the customer by using the time-based preprocessing data as shown in FIG. 2 and predicts the departure of the customer.

The behavior pattern extraction model 310 may be a predictive model based on RNN (Recurrent Neural Network), a convolutional neural network (CNN) based prediction model.

RNN is a deep learning model for learning data that changes over time, such as time-series data, and is an artificial neural network constructed by connecting a network at a reference time point t and a next time point t + 1. (ANN).

LSTM is proposed to overcome the vanishing gradient problem that occurs in RNN. LSTM slows down the dilution of information by controlling the amount of information flowing into each node using gates. Apply input gate and forget gate to cell state to selectively add or remove information. By opening and closing the gate, the degree of dilution of hidden node information can be controlled to prevent vanishing gradient. LSTM is a kind of RNN.

CNN is a kind of multi-layer neural network and shows good performance in pattern recognition field. It is mainly used to classify and recognize data such as image and text. The main elements of CNN are the convolution layer and the poolng layer. Features of the original data are extracted through a convolution layer that extracts the features of the input and a pooling layer that reduces the input without losing important information.

As shown in FIGS. 4 and 5, the behavior pattern extraction model 310 constructs an RNN model (FIG. 4) or a CNN model that extracts a behavior pattern of a customer using time-based preprocessing data (TD). Concatenate operation pattern of customer's behavior extracted by each model and apply to artificial neural network. In other words, predict customer churn by independently using customer behavior patterns broken down by criteria attributes.

The behavior pattern extraction model 310 may consist of any one of an RNN model and a CNN model, or may consist of an RNN model and a CNN model.

The statistical index extraction model 320 extracts the statistical index of the customer using the customer-based preprocessing data as shown in FIG. 3 to predict whether the customer has left.

The statistical index extraction model 320 may be formed of at least one of an RF (Random Forest, Random Forest) based prediction model and an XGBoost (Extreme Gradient Boosting) based prediction model, and more specifically, an RF, XGBoost based prediction model. In addition, it is desirable to build a statistical index extraction model 320 together.

RF is a decision tree based machine learning technique, which is an ensemble technique that combines several models of low performance (low accuracy) to form a high performance model. To perform the classification, each tree “votes” for a particular class and classifies the most votes into one class. It has the advantage of being able to calculate the importance of independent variables and reduce overfitting.

XGBoost is a highly efficient and scalable machine learning method for tree boosting. It is used in various fields because it shows higher performance than the existing tree boosting method. As one of the tree ensemble techniques, the parallel processing technique is used to construct the tree, which has the advantage of fast execution time.

When the statistical index extraction model 320 is constructed with an RF and an XGBoost-based prediction model, first, the importance of each attribute included in the customer-based preprocessing data is calculated using the RF. The importance of the property is calculated by calculating the Mean Decrease Accuracy. When the Random Forest is started, it is calculated by the algorithm inside the Random Forest. As a result of importance calculation of each property, RF model and XGBoost model are constructed by selecting a predetermined number of properties with high importance. Here, the predetermined number may be 100, for example.

In other words, instead of building the RF model and the XGBoost model using all the attributes included in the customer-based preprocessing data, the top 100 attributes are selected from the attributes calculated by the RF model and arranged in order of importance. Construct the RF model and the XGBoost model using the attributes.

The ensemble prediction model 330 applies a voting technique to the prediction results of the behavior pattern extraction model 310 and the statistical index extraction model 320 to calculate more than half of the prediction results. More specifically, the ensemble prediction model 330 may be constructed of an RNN model using a customer's behavior pattern, an RNN model with high accuracy among CNN models, and an RF model and an XGBoost model using a customer's statistical indicators. That is, the ensemble prediction model 330 may be constructed as an RNN model, an RF model, and an XGBoost model. Here, the ensemble prediction model 330 does not rebuild the RNN model, the RF model, or the XGBoost model, but rather the behavior pattern. The ensemble prediction model 330 is constructed by using the RNN model, the RF model, and the XGBoost model used for the extraction model 310 and the statistical index extraction model 320.

The estimating prediction process of the ensemble prediction model 330 may include at least two or more results of predicting the deviation of the RNN model, the RF model, and the XGBoost model, which construct the behavior pattern extraction model 310 and the statistical index extraction model 320. If it predicts this "deviation" and the other predicts it as "non-deviation", the ensemble prediction model 330 predicts the majority "violation" according to the voting technique.

The output unit 400 outputs the prediction result of the customer deviation predictor 300 to the outside through visual or audio means.

Next, a customer departure prediction method according to an embodiment of the present invention will be described.

The customer departure prediction method according to the embodiment of the present invention may be performed by a computing means. Here, the computing means means a terminal having a memory for storing a program such as a desktop computer, a notebook computer, a tablet PC, a mobile communication terminal, a microprocessor for executing a program and controlling the same. Such computing means may include an input unit 100, a preprocessor 200, a customer departure predictor 300, an output unit 400, or the like, which are components of the above-described customer deviation predictor, or execute their functions. A program stored in the recording medium may be mounted to perform a customer departure prediction method.

The customer departure prediction method according to the embodiment of the present invention comprises a first step of inputting customer data, a second step of pre-processing the input customer data and classifying the time-based preprocessing data and the customer-based preprocessing data, and the classified time. A third step of extracting behavior patterns using the standard preprocessing data, extracting customer statistical indicators using the classified customer standard preprocessing data, and using the extracted behavior patterns and statistical indicators to predict whether the customer has left The fourth step of outputting the prediction result of the customer deviation prediction unit.

The computing means preprocesses the customer data input in the first step on a time basis and classifies the data into time based preprocessing data. In addition, the customer data input in the first step is pre-processed on a customer basis and classified into customer-based preprocessing data.

Here, the time-based preprocessing data may be in a time table format in which customer behaviors are arranged in chronological order. In addition, the customer-based preprocessing data may be in an attribute table format in which a plurality of customer attributes are arranged regardless of time.

In the third step of estimating the departure of the customer, a process of estimating the departure of the customer by extracting a behavior pattern using the classified time-based preprocessing data and extracting statistical indicators using the classified customer-based preprocessing data It includes the process of predicting whether the customer has deviated, and the process of calculating the majority of the prediction result by applying the voting technique to the prediction result.

Behavior patterns can be extracted using RNN-based prediction models and CNN-based prediction models. In addition, statistical indicators can be extracted using an RF-based prediction model or an XGBoost-based prediction model.

In the third step, at least any three prediction models of RNN, CNN, RF, and XGBoost based prediction models may be used, and the process of calculating a majority or more prediction results by applying a voting technique to the prediction results may be performed based on RNN. It is preferable to use the prediction model, the RF-based prediction model, and the XGBoost-based prediction model.

Next, look at the improved accuracy of the customer departure prediction apparatus according to an embodiment of the present invention through the experimental data.

1. Experimental data

This experiment used log data of game Blade & Soul. It consists of 4000 training data, 3000 validation data and 3000 test data, and one data is log data collected for 8 weeks for one customer.

The data used in this experiment was provided by NCsoft and was released at 2017 Game Data Mining competition (IEEE Computational Intelligence in Games). The configuration of the data is shown in the table shown in FIG.

In Figure 6, "churn" means that the departure. Model is created through training data, reliability is obtained through validation data, and final accuracy is calculated using test data. Since the number of customers who deviated and the number of customers who did not deviate was unbalanced, Oversampling was applied to solve the reduction in model performance due to data imbalance.

5.2 building predictive models

We used the character that the customer plays as the reference attribute for the time base preprocessing. In the game used in this experiment, one user can own several characters. Therefore, the character with the longest play time was defined as the main character and the remaining characters as auxiliary characters, so that the behavior pattern of one customer was divided into the behavior pattern of two characters.

Construct RNN prediction model using LSTM. The RNN prediction model consists of five layers, and each layer consists of 56 LSTM cells.

The CNN prediction model consists of five layers, with one layer consisting of 3 × 1 convolution and 3 × 1 max pooling.

The RNN prediction model and the CNN prediction model used Exponential Linear Units (ELU) (1.0) as activation functions. Loss uses cross entropy and Root Mean Square Propagation (RMSProp) as an optimizer. In addition, Batch normalization and Dropout (0.4) were applied to each layer.

In order to construct the RF prediction model and the XGBoost prediction model, 1364 attributes (statistical indicators) representing the statistical characteristics of each customer and the behavioral patterns (behavior patterns) reflecting the change of the customer's behavior were generated. Mean Decrease Accuracy of each attribute was calculated using RF to select the attribute to be used for constructing the prediction model. Based on the calculated importance, 100 attributes were selected to build an RF prediction model and an XGBoost prediction model.

3. Experimental Results

The accuracy of the prediction model is shown in the table shown in FIG. Among the predictive models using one feature, the model with the highest accuracy is the predictive model that extracts the behavior patterns of customers using RNN. The second highest-accuracy model is the XGBoost prediction model that utilizes the customer's statistical characteristics (statistical indicators).

The accuracy of the ensemble prediction model constructed using the RNN model and the XGBoost model was the highest at 76%. In other words, the performance of the model that predicts customer churn by using both customer behavior patterns and customer statistical characteristics (statistic indicators) was the best.

In addition, the RNN model and the CNN model were constructed using undivided data using the reference attributes. Through this, we verified whether segmentation of customer behavior patterns using reference attributes has an effective effect on predicting customer churn.

In addition, the RF model and the XGBoost model were constructed using data that did not include behavior change reflection attributes. Through this, we confirmed whether the behavior change reflecting effect has an effect on the prediction of customer churn. Experimental results are shown in the table shown in FIG. 8.

The accuracy of the RNN and CNN models, which extracted the behavioral patterns by dividing the behavioral data of the customer by using the reference attribute, was 3.5% higher than the accuracy of the models that did not. Therefore, it was found that segmenting and extracting customer's behavior patterns according to the criteria attributes is effective for predicting customer churn.

The accuracy of the XGBoost and RF models, using attributes that reflect changes in customer behavior, averaged 6.5% higher than those of other models. In addition, the importance (Mean Decrease Accuracy) of each property calculated to build the XGBoost model and RF model is shown in FIG.

The most important attribute is the amount of experience gained each week by the customer. The second most important attribute is the number of times the customer converted items in the last week. The third most important attribute is the coefficient of variation in weekly play time. The top three attributes reflect changes in customer behavior. Therefore, it can be seen that it is effective to include an attribute reflecting the change in the behavior of the customer when predicting the customer departure using the statistical characteristics of the customer.

According to the customer departure predicting apparatus according to the embodiment of the present invention as described above, it is possible to determine whether the customer leaving with more improved accuracy.

As mentioned above, although an embodiment of the present invention has been described, those of ordinary skill in the art may add, change, delete or add components within the scope not departing from the spirit of the present invention described in the claims. The present invention may be modified and changed in various ways, etc., which will also be included within the scope of the present invention.

100 input unit 200 preprocessing unit
210: preprocessing module based on time 220: preprocessing module based on customers
300: customer deviation prediction unit 310: behavior pattern extraction model
320: Statistical Indicator Extraction Model 330: Ensemble Prediction Model
400: output unit

Claims (15)

An input unit to which customer data is input;
A pre-processing unit for pre-processing the customer data input to the input unit and classifying the time-based pre-processing data and the customer-based pre-processing data;
The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting customer churn prediction unit; And,
Output unit for outputting the prediction result of the customer deviation prediction unit
Customer departure prediction apparatus comprising a.
The method according to claim 1, wherein the preprocessing unit,
A time-based preprocessing module for pre-processing customer data input to the input unit on a time-based basis and classifying the data into time-based pre-processing data;
Customer-based preprocessing module for pre-processing customer data input to the input unit on a customer basis and classifying it as customer-based preprocessing data
Customer departure prediction apparatus comprising a.
The method according to claim 2,
Wherein the time-based preprocessing data is a time table format in which customer behaviors are arranged in chronological order, and the customer-based preprocessing data is an attribute table format in which a plurality of customer attributes are arranged regardless of time.
The method of claim 1, wherein the customer departure prediction unit,
A behavior pattern extraction model that predicts a customer departure by extracting a behavior pattern using the classified time based preprocessing data;
A statistical index extraction model for predicting whether the customer deviates by extracting the statistical index using the classified customer standard preprocessing data;
An ensemble prediction model that produces a majority or more prediction results by applying a voting technique to the prediction results of the behavior pattern extraction model and the statistical index extraction model.
Customer departure prediction apparatus comprising a.
The method according to claim 4,
The behavior pattern extraction model is a predictive model based on RNN (Recurrent Neural Network), a convolutional neural network (CNN) based prediction model,
The statistical indicator extraction model is a customer departure prediction device which is a prediction model based on a random forest (RF) or a prediction model based on XGBoost (Extreme Gradient Boosting).
The method of claim 5, wherein the customer departure prediction unit,
The customer churn prediction apparatus consisting of at least three prediction models of the RNN, CNN, RF, XGBoost-based prediction model.
The method according to claim 5, wherein the ensemble prediction model,
The customer departure prediction device comprising the RNN-based prediction model, the RF-based prediction model, and the XGBoost-based prediction model.
Performed by computing means,
A first step of entering customer data;
A second step of pre-processing the input customer data and classifying it into time-based preprocessing data and customer-based preprocessing data;
The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting a third step; And,
A fourth step of outputting a prediction result of the customer deviation prediction unit;
Customer deviation prediction method comprising a.
The method of claim 8, wherein the second step,
Pre-processing the input customer data on a time basis to classify the time-based preprocessing data, and pre-processing the input customer data on a customer basis to classify the customer departure pre-processing data.
The method according to claim 9,
The time-based preprocessing data is a time table format in which customer behaviors are arranged in chronological order, and the customer-based preprocessing data is an attribute table format in which a plurality of customer attributes are arranged regardless of time.
The method of claim 8, wherein the third step,
Extracting behavior patterns using the classified time-based preprocessing data to predict whether the customer has left, and extracting statistical indicators using the classified customer-based preprocessing data to predict whether the customer has left; A method of predicting customer churn comprising applying a voting technique to the prediction result to produce a majority or more prediction result.
The method according to claim 11,
The behavior pattern is extracted using a RNN based prediction model, a CNN based prediction model,
The statistical indicator is a customer departure prediction method to extract using the RF-based or XGBoost-based prediction model.
The method of claim 12, wherein the third step,
The customer churn prediction method using at least any three prediction models of the RNN, CNN, RF, XGBoost-based prediction model.
The method according to claim 12,
The process of calculating a majority or more prediction results by applying a voting technique to the prediction results,
A customer departure prediction method using the RNN based prediction model, the RF based prediction model, and the XGBoost based prediction model.
In computing means,
Entering customer data;
Pre-processing the input customer data and classifying the received customer data into time-based preprocessing data and customer-based preprocessing data;
The behavioral pattern is extracted using the classified time based preprocessing data, the statistical index of the customer is extracted using the classified customer based preprocessing data, and whether the customer is separated using the extracted behavior pattern and the statistical index. Predicting; And,
Outputting a prediction result of the customer deviation prediction unit;
A computer program stored in a recording medium for executing the.

Images