KR102475513B1 - Method and system for automatically generating customer lifetime value (CLV) prediction models for each stage of customer life - Google Patents

Abstract

The present invention relates to a method and system for automatically generating a customer lifetime value (CLV) prediction model by stage of a customer lifecycle. The method for automatically generating a customer lifetime value (CLV) prediction model by stage of a customer lifecycle may comprise the steps of: collecting customer information comprising at least purchase history information of a customer; calculating a sum of purchase amounts (CV) comprised in the customer's purchase history information and classifying life cycle stages of the customer based on the calculated sum of purchase amounts and the number of purchases; generating a learning model (M) for predicting future customer value with customer data for each life cycle stage and prediction cycle; calculating the customer's realized CV value after a preset unit of time; and learning the model for each life cycle stage and prediction cycle of the customer. According to the present invention, the customer value prediction model can be updated in real time, and the value prediction model for each customer level can be simultaneously learned to increase the value prediction accuracy. The influence can be reflected in the changes in marketing activities and purchase trends, and the performance of the learning model can be prevented from deteriorating. In addition, the size of the data set (time window) to be trained periodically can be adjusted. Thus, the cyclical nature of changes in the external environment can be reflected.

Description

Method and system for automatically generating customer lifetime value (CLV) prediction models for each stage of customer life}

The present invention classifies the customer's life stage and automates the creation of a machine learning model for each life stage by using AutoML technology, thereby predicting customer value by stage of the customer's life stage, which can update the customer value prediction model in real time. It relates to methods and systems.

Customer Lifetime Value (CLV) is also known as customer lifetime value or LTV. CLV is an important index for evaluating the financial value of customers. Referring to the Pareto law that 20% of customers generate the most sales, it is used as an index to find customers who contribute the most to corporate sales and manage good customers. .

The CLV model can be expressed as Equation 1.

[Equation 1]

Here, M is the average sales per customer, usually calculated per year, c is the average cost per customer, usually calculated per year, and r is the customer retention rate, which customer The probability of remaining a customer in the following year, where i is the interest rate or discount rate, and AC is the acquisition cost, which is the cost of getting a customer to make their first visit or first purchase.

To increase customer lifetime value, you need to keep customers buying as much as possible and make them buy more often, a lot, and for a longer period of time.

Prediction of customer lifetime value is important data for targeting customers for marketing. However, existing customer lifecycle prediction methods lacked consideration for prediction by customer lifecycle stage. Customers who are not loyal customers at all stages, such as awareness, registration, and purchase, can identify customers who are likely to go to the loyalty stage first, in that effective marketing is possible. It is necessary to simultaneously train the value prediction model to increase the accuracy of value prediction. Customers are influenced by marketing activities and are affected by changes in purchasing trends. Accordingly, if marketing activities and trend changes are not continuously reflected, the performance of the existing learned customer value prediction model deteriorates, so continuous model learning is required.

Registered Patent Publication No. 10-2237399, 2021.04.01)

The problem to be solved by the present invention has been created to solve the above-mentioned problems, while generating and learning a customer life cycle value prediction model and a customer life stage change prediction model for each customer stage, and analyzing customer attribute information and customer purchasing behavior analysis result information. To provide a method and system for automatically generating and learning a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle, which automatically creates and learns an ML model periodically or continuously when data changes occur through AutoML technology.

A method for automatically generating a customer lifetime value (CLV) prediction model according to an aspect of the present invention for achieving the above technical problem includes collecting customer information including at least customer purchase history information; calculating a sum of purchase amounts (CV) included in the customer's purchase history information and classifying life cycle stages of the customer based on the calculated sum of purchase amounts and the number of purchases; generating a learning model (M) for predicting future customer value with customer data for each life cycle stage and prediction cycle; calculating the customer's realized CV value after a preset unit of time; and learning the model for each life cycle stage and prediction cycle of the customer.

The customer information may include at least one of attribute information including age and region of the customer, number of store or site visits, shopping cart conversion rate, behavior information including purchase history, and opinion information including comments.

The life cycle stages of the customers include new customers without purchases, first-time purchasers who make purchases for the first time, regular customers whose number of purchases is less than the first threshold, regular customers whose number of purchases is greater than the first threshold, and customer lifetime value ( CLV) may include loyal customers of a certain size or higher.

The learning model (M) determines a specific point in time from customer information up to now, uses data before the specific point in time as features, and uses customer value (purchase amount) as a label from data after the specific point in time to determine customer lifetime value. Predictable.

In the step of learning the model, the model may be learned for each stage of the customer's life cycle and for each prediction period in a sliding time window manner.

According to one aspect of the present invention for achieving the above technical problem, the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle includes a customer information collection unit that collects customer information including at least purchase history information of the customer; a life cycle classification unit that calculates a sum of purchase amounts (CV) included in the customer purchase history information and classifies life cycle stages of the customer based on the calculated sum of purchase amounts and the number of purchases; an AUTOML unit for generating a learning model (M) for predicting future customer value with customer data for each life cycle stage and prediction cycle; a model learning unit that learns a model for each customer's life cycle stage and prediction cycle; and a CV calculation unit that calculates a realized CV value of the customer after a preset unit time has elapsed.

The customer information collection unit may collect at least one of attribute information including age and region of the customer, number of visits to a store or site, shopping cart conversion rate, behavior information including purchase history, and opinion information including comments. The customer life cycle classification unit classifies customers into new customers without purchases, first-time purchasers who purchase for the first time, regular customers whose number of purchases is less than the first threshold, regular customers whose number of purchases is greater than the first threshold, and customer life cycles. It can be classified as a loyal customer whose value (CLV) exceeds a certain level.

The learning model (M) in the AUTOML unit determines a specific point in time from customer information up to now, sets data before the specific point in time as features, and labels customer value (purchase amount) from data after the specific point in time. You can predict customer lifetime value. The model learning unit may learn a model for each stage of a customer's life cycle and for each prediction cycle in a sliding time window manner.

According to the method and system for automatically generating and learning a customer lifetime value (CLV) prediction model for each stage of a customer's life according to the present invention, by classifying a customer's life stage and automating a machine learning (ML) model for each life stage using AutoML technology, the customer The value prediction model can be updated in real time, and the value prediction accuracy can be increased by simultaneously learning the value prediction model for each customer stage.

In addition, according to the present invention, by allowing the ML model to learn periodically or continuously when data changes for each customer stage, it is possible to reflect the influence of changes in marketing activities and purchase trends, and to prevent the performance of the learning model from deteriorating. have. In addition, it is possible to provide a customer lifetime value prediction model that reflects changes in the external environment by establishing an automatic learning system for the same data periodically or when an event occurs. In particular, in the present invention, by adjusting the size (time window) of the data set to be periodically learned, it is possible to reflect the periodic characteristics of the change in the external environment.

In addition, according to the present invention, efficient predictive marketing can be performed by utilizing various information such as customer attributes, behavioral log data, and external environment as predictive factors.

1 is a block diagram showing the configuration of an embodiment of a system for automatically generating a customer lifetime value (CLV) prediction model by stage of the customer life according to the present invention.
2 is a feature of the data prior to a specific point in the AUTOML unit of the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of the customer life cycle according to an embodiment of the present invention, and from the data after the specific point in time It shows a conceptual diagram that predicts customer value (CV) using customer value (purchase amount) as a label.
3 shows an example of calculating customer value using the size of a circle proportional to the size of the purchase amount in the CV calculator of the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of the customer life according to an embodiment of the present invention. will be.
4 illustrates learning of a predictive model through a sliding time window in the model learning unit of the system for automatically generating a customer lifetime value (CLV) predictive model for each stage of a customer lifecycle according to an embodiment of the present invention.
5 shows that the size of the sliding time window and the period dividing (Data, CV) to be used for learning may vary in the model for each prediction period in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of the customer life cycle according to an embodiment of the present invention. indicates that there is
Figure 6 shows the classification of customer life cycle stages for predicting loyal customers in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention.
7 illustrates classifying customer stages based on the number of purchases and purchase amount in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's life cycle according to an embodiment of the present invention.
8 is a system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's life cycle according to an embodiment of the present invention, by learning a customer value prediction model for each stage of a customer's life cycle, so that a CV reflecting future value can become a loyal customer. It is a conceptual diagram that explains
9 is a diagram illustrating the number of predictive models to be trained in the system for automatically generating a customer lifetime value (CLV) predictive model for each stage of a customer lifecycle according to an embodiment of the present invention.
10 is a flowchart illustrating a method of automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention.
11 shows the range of AutoML used to predict customer lifetime value (CLV) by stage of customer lifecycle.
12 shows the range of AutoML for continuous learning to predict the customer lifetime value (CLV) for each stage of the customer lifecycle according to the present invention.
13 shows that the size of the sliding time window can be increased or decreased in order to generate an optimal prediction model when learning each model for predicting the customer lifetime value (CLV) at each stage of the customer lifecycle according to the present invention.
14 shows an example of data in which the system for automatically generating a customer lifetime value (CLV) prediction model by stage according to the present invention is composed of data, a prediction model (M), and a predicted customer value (predicted CV).
15 shows the overall characteristics of the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Since the embodiments described in this specification and the configurations shown in the drawings are only one preferred embodiment of the present invention and do not represent all of the technical spirit of the present invention, various equivalents that can replace them at the time of the present application It should be understood that there may be variations and variations.

The present invention continuously automatically creates and learns a machine learning (ML) model by using AutoML technology for a customer life cycle value prediction model and a customer life stage change prediction model for each customer stage. In addition, ML models are learned using customer attribute information and customer purchasing behavior analysis result information, and ML models are continuously learned periodically or when data changes occur using AutoML technology.

1 is a block diagram showing the configuration of an embodiment of a system for automatically generating a customer lifetime value (CLV) prediction model by stage of the customer life according to the present invention. According to an embodiment of the present invention, the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of the customer life cycle includes a customer information collection unit 110, a life cycle classification unit 120, an AUTOML unit 130, a CV calculation unit ( 140) and a model learning unit 150.

The customer information collection unit 110 collects customer information including at least purchase history information of the customer. In the present invention, the machine learning (ML) model may utilize various types of customer information in addition to purchase history. The customer information collection unit 110 collects customer attribute information including age and region of the customer, number of store or site visits, shopping cart conversion rate, customer behavior information including purchase history, customer opinion information including comments, and the like. can be collected

The life cycle classification unit 120 calculates the sum of purchase amounts (CV) included in the customer purchase history information and classifies the life cycle stages of the customer based on the calculated sum of purchase amounts and the number of purchases.

There can be many ways to classify customers based on Recency, Frequency, and Monetary. That is, there may be various RFM-based customer stage divisions. The customer life cycle stage used in the present invention is, for example, the life cycle classification unit 120 of the customer is a new customer who has not purchased a customer, a first purchase customer who has made a purchase for the first time, and a general customer whose number of purchases is less than the first threshold value. Customers, regular customers whose number of purchases is greater than the first threshold, and loyal customers whose customer lifetime value (CLV) is greater than or equal to a certain level may be classified.

The present invention can create a predictive model for each stage of the customer life cycle, automate the model generation, predict the probability of reaching the final stage (loyal customer) from the front stage of the customer, and use it for marketing. If this is applied to manufacturing, it is possible to prepare for an increase or decrease in yield by predicting yield for each process step, and through this, it is possible to increase inventory or prepare sales activities in advance.

The AUTOML unit 130 generates a learning model (M) for predicting future customer value with customer data for each life cycle stage and prediction cycle. As shown in FIG. 2, the learning model (M) determines a specific point in time from customer information up to now, sets the data before the specific point in time as features, and calculates the customer value from the data after the specific point in time. Using (purchase amount) as a label, customer value (CV) can be predicted. Provides a customer value (CV) prediction model based on past attribute information, behavioral information, and opinion information of different customers for each customer.

Learning model M is a model that predicts future customer value using customer data, and learns a model that predicts customer value (CV) with data. Model M returns a predicted future customer value (CV) given the data of the i-th customer (Ci).

The CV calculator 140 calculates a customer's realized customer value (CV) value after a preset unit time elapses. The method of calculating customer value in the present invention can be made as follows. Each customer has a purchase history. A common approach is to predict future customer value based on purchase history prior to the current point in time, and this is done using a probabilistic model. Customer value is the amount of revenue contributed by customers during a given period. In general, CLV (Customer Lifetime Value) refers to the financial value a customer gives to a company over their lifetime, and refers to the total sales contribution from the first purchase to the point of no further purchases. Sales generated by customers can be divided into already realized sales and predicted sales for each customer, and correspondingly, there are realized customer value (CV) and predicted customer value. 3 shows an example showing customer value calculation. Referring to FIG. 3 , based on the current point in time t _now , customer 1 has made purchases four times, and the sizes of blue circles 310 , 315 , 320 , and 325 indicate the size of the purchase amount. Further, the dotted line circles 330 and 335 represent the predicted customer value, and the size of the dotted circle represents the size of the predicted customer value. Likewise, customer 2 has made purchases three times (340, 345, 350), and the dotted line circle 360 represents the predicted customer value based on this.

The model learning unit 150 learns a model for each stage of the customer's life cycle and each prediction cycle, and can learn a model for each stage of the customer's life cycle and each prediction cycle in a sliding time window manner.

The present invention continuously learns a predictive model through a sliding time window. 4 shows learning a predictive model through a sliding time window. Referring to FIG. 4, if the current time is t _now , there is an M _prev model learned before t _now , and the M _prev model is a prediction model learned from previous data (Data _prev ) and previous customer value (CV _prev ) , and the previous value forecast for each customer is the predicted CV _prev .

Continuous learning continues to train the model periodically or whenever a specific event occurs. The learning target data has a time window and utilizes data within a given time range.

The actual calculated CV value after t _prev may differ from the previous predicted value predicted CV _prev , which reflects the prediction error value of the M _prev model and changes in the external environment or purchase trend from t _prev to period t. It can be viewed as the sum of data drift values.

When learning a new predictive model M, the predictive model M can be created by inheriting M _prev and learning it with a new training data set (Data, CV). Also, to preserve the knowledge learned in the previous model, | CV - PredictionCV _prev | Only customers with < CV _Threshold can be selected and M can be trained. Through the CV _Threshold value, it is possible to control the allowable error range between the predicted value and the actual value.

According to the embodiment of the present invention, since it is predicted by period and a sliding time window method is used, in RFM-based customer stage classification, 　recency is already reflected in the model, so the number of purchases (Frequency) and purchase amount ( Moneraty) can be the center of the model.

A model for each prediction period in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention will be described. When training a model in one step, you can construct short-term and long-term predictive models separately. Depending on the marketing method, short-term/long-term value predictions may affect target customer selection. Long-term forecasts or short-term forecasts can be used, depending on whether the marketing is for long-term benefit loyalty marketing for VIP customers or marketing to increase sales in a short period of time.

5 shows that the size of the sliding time window and the period dividing (Data, CV) to be used for learning may vary in the model for each prediction period in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of the customer life cycle according to an embodiment of the present invention. indicates that there is Referring to FIG. 5, the present invention sets the prediction period to 1 week, 1 month, 3 months, 1 year, and customer lifetime during model learning. The model can be divided into parts. Accordingly, the size of the sliding time window and the period dividing (Data, CV) to be used for learning may vary.

According to the present invention, the classification of customer life cycle stages for predicting loyal customers will be described. An important part of corporate marketing is to predict loyal customers in advance from the customer inflow stage. From the perspective of future customer value, it is possible to classify the stages of the customer's life cycle. Figure 6 shows the classification of customer life cycle stages for predicting loyal customers in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention. Referring to FIG. 6, for example, the life cycle of a customer is divided into new customers without purchase (610), first-time purchase customers (620), regular customers (630), regular customers (640), and loyal customers (650). It can be done in stages.

In addition, each customer's stage can be divided based on the number of purchases and the purchase amount. 7 illustrates classifying customer stages based on the number of purchases and purchase amount in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's life cycle according to an embodiment of the present invention. Referring to FIG. 7, if the customer value (CV) is above a certain size, loyal customers (CV_loyal, 750), if the number of purchases is above a certain size, regular customers (Frequency, 740), first-time purchase customers (720), and regular customers (730), etc., and can be defined according to the number of purchases or purchase amount. Possible marketing goals for each step to increase customer value may be different. For example, marketing may be performed so that a general customer 730 may become a regular customer by identifying a customer purchase intention and recommending various products.

In the present invention, by learning a customer value prediction model for each customer's life cycle stage, it is possible to identify customers whose CV reflecting future value is a loyal customer. 8 is a system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's life cycle according to an embodiment of the present invention, by learning a customer value prediction model for each stage of a customer's life cycle, so that a CV reflecting future value can become a loyal customer. It is a conceptual diagram that explains

Meanwhile, the number of prediction models to be trained in the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life according to an embodiment according to the present invention will be described. 9 is a diagram illustrating the number of predictive models to be trained in the system for automatically generating a customer lifetime value (CLV) predictive model for each stage of a customer lifecycle according to an embodiment of the present invention. Referring to FIG. 9 , the number of prediction models to be trained at one time is the product of the number of customer life cycle stages and the number of prediction period types. For example, you can compute an updated customer value once a week, compare the predictions of previous models and reflect the differences. For example, if a year is 52 weeks, there are 5 customer life cycles (e.g. new, first purchase, regular, regular, loyalty) and 5 forecast periods (e.g. one week, one month, three months, one year, infinite ( No period limitation)) on a weekly basis, the number of prediction models to be trained is 1,300 (= 52x5x5) models.

10 is a flowchart illustrating a method of automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention. Referring to FIGS. 1 and 10 , a method for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to an embodiment of the present invention will be described.

First, the customer information collection unit 110 collects customer information including at least the customer's purchase history information (step S1010). The customer information includes the customer's age, attribute information including region, number of visits to stores or sites, and shopping cart. At least one of behavioral information including a purchase conversion rate, purchase history, and opinion information including comments may be included. That is, customer information may include customer attribute information, customer behavior information, and customer opinion information. The customer information needs to include the customer's purchase history information for CV calculation. The purchase history information may include the customer's purchase date, purchase amount information, and the like. Information other than purchase history information is used as data to predict CV. (Data, CV)

When customer information is collected, the life cycle classification unit 120 calculates the sum of purchase amounts (CV) included in the customer purchase history information and classifies the life cycle stages of the customer based on the calculated sum of purchase amounts and the number of purchases. (Step S1020) The stages of the customer's life cycle include new customers without purchases, first-time purchasers who purchase for the first time, regular customers whose number of purchases is less than the first threshold, and regular customers whose number of purchases is greater than the first threshold. It may include customers and loyal customers whose customer lifetime value (CLV) exceeds a certain level.

That is, when customer information is collected, the realized CV is calculated, and based on this, the life cycle classification unit 120 classifies the customer's life cycle stage. The realized CV is calculated as the sum of the purchase amount from the customer purchase history information used for learning, and the customer's life cycle stage is classified according to the customer's total CV and the number of purchases. For first-time customers, purchases may not be included within this period. Whether to judge the customer as a churn customer or as a first-time purchaser is determined as needed. Loyal customers and other customers are classified according to the threshold of total purchases, and the rest of the regular customers and regular customers are classified according to the threshold of the number of purchases. A general customer can be classified as a customer with a purchase history of two or more times.

The AUTOML unit 130 generates a learning model (M) for predicting future customer value with customer data for each life cycle stage and prediction cycle (step S1030). The learning model (M) is the customer information up to now. It is possible to predict a customer lifetime value by setting a specific point in time and using data before the specific point in time as features and using customer value (purchase amount) as a label from data after the specific point in time. If there is no previous customer value prediction model M_0, model M_0 can be constructed with (Data, CV), which is a set of data and customer value.

The CV calculation unit 140 calculates the realized CV value of the customer after a preset unit time has elapsed (step S1040). CV values can be calculated.

The model learning unit 150 learns the model for each stage of the customer's life cycle and each prediction cycle. (Step S1050) The model learning may be performed for each stage of the customer's life cycle and each prediction cycle in a sliding time window method.

By repeatedly performing steps S1040 and S1050, it is possible to predict customer lifetime value (CLV) for each stage of a customer's lifecycle.

Meanwhile, AutoML predicting the customer lifetime value (CLV) for each stage of the customer lifecycle according to the present invention will be described in more detail. Manual learning by analysts is impossible to periodically train the model for each customer lifecycle stage by forecast period.

Automated machine learning (AutoML) is the process of automating manual and repetitive tasks in the machine learning pipeline, automating the time-consuming coding, algorithm selection, and tuning tasks of developing predictive models. AutoML enables machine learning (ML) learning to be automated by generating optimal classification and prediction models given data for learning.

11 shows the range of AutoML used to predict customer lifetime value (CLV) by stage of customer lifecycle. AutoML can be composed of the process of feature engineering, model (algorithm) selection, and hyper parameter optimization. (There is also AutoML technology that includes a data preparation process that compensates for similar features or missing features, but in the present invention, although there may be some variations, data preparation process is excluded from the scope of AutoML because it continuously learns from the same data set. to do).

Feature engineering is a process of constructing parameters necessary for learning from data, and is a process of creating optimal model performance by creating new features or removing features that do not affect the model.

Algorithm selection is the process of reviewing and selecting various algorithms to create optimal classification and prediction models from given data. For prediction models, algorithms such as linear regression and logistic regression, and for classification problems, known algorithms such as random forest, Bayesian, and support vector machines are set as candidates, and various experiments are automatically executed.

Hyperparameter optimization is the process of optimizing hyperparameters in a selected algorithm. There are two types of parameters: model parameters and hyperparameters. The slope of linear regression or the value of the connection strength between neurons in deep learning are model parameters obtained through learning, and the number of trees or maximum depth in a random forest are set before learning. It is a hyperparameter. Hyperparameters in neural networks include a learning rate and the number of times of learning.

12 shows the range of AutoML for continuous learning to predict the customer lifetime value (CLV) for each stage of the customer lifecycle according to the present invention. Recently, various AutoML open source and commercial products have been released and used. There are many data analysis tasks to be solved, but it can be used to improve analysis efficiency in companies that lack data analyst manpower. For commercial use, there are MS Azure ML, Google Cloud AutoML, AWS SageMaker, DataRobot, etc., and for open source, there are H2O, Auto-sklearn, TPOT, etc.

In the present invention, AutoML is used for model learning, and existing AutoML is used to perform unit tasks to create models using data given by AutoML, but is not used as an infrastructure for continuous learning. A core part of the present invention is a customer lifetime value prediction model that reflects external environment changes by establishing an automatic learning system for the same data periodically or when an event occurs.

Furthermore, in the present invention, by adjusting the size (time window) of the data set to be periodically learned, the periodic characteristics of the change in the external environment are reflected.

As shown in FIG. 13, an optimal prediction model is created by changing the size of the sliding time window during each model training, and this time window can be increased or decreased during each learning.

14 shows that the system for automatically generating a customer lifetime value (CLV) prediction model by stage of the customer life according to the present invention is composed of Data (1410), a prediction model (M, 1420), and a predicted customer value (predicted CV, 1430), An example of the data is shown. Referring to FIG. 14 , the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's life cycle according to an embodiment of the present invention accepts demographic information, purchase behavior information, and opinion information as data 1410 and generates customer values. Predictive models can predict customer value. Demographic information may include age, gender, and region, purchase behavior information may include the number of visits, shopping cart, purchase history, recent purchase date, total number of purchases, total purchase amount, etc., and opinion information may include the number of participants, emotional index may be included.

Meanwhile, business improvement scenarios based on CLV applicable to the present invention are as follows.

1) Cart, Browse, and Search Abandonment

A lot of marketing costs have already been spent on finding products and putting them in the shopping cart by customers who have been brought in through site advertisements. In fact, 69% of online shopping carts abandon purchases when items in the shopping cart are abandoned, which is ineffective marketing.

Retargeting and sending cart cancel emails are necessary to increase site visits and click-through rates and increase sales and profits. Marketing considering the customer life stage is to secure new customers and regular/loyal customers, so personalized marketing is needed to inform discount information on the product and show products purchased together.

2) Onboarding and Early Activation

For new customers, it is most important to quickly judge the customer's intention or taste in the early stage. Activities such as sending special discount coupons for products on the search or cart list to customers who have signed up and have not ordered for a certain period of time are required.

3) Upselling and Cross-Selling

Introduce products with better quality and functions to increase the purchase amount or suggest complementary products to induce additional purchases. Encourage additional visits from customers and improve CLV by recommending similar products of other functions or brands before purchasing a product, or by presenting other products that are needed together during or after purchase.

4) Segmentation and Drip Campaigns

Automatically sending marketing emails (Drip Campaigns) according to customer segments or according to a schedule tailored to customer tastes can be an effective means of inducing customers to take interest in a product.

And, knowing CLV, you can establish an appropriate budget to secure new customers. The cost of acquiring new customers is called CPA (Cost Per Acquisition). Without measuring or estimating the average cost of a customer's purchase, it is difficult to determine how much to spend on customer acquisition marketing. If you don't have a CLV value, you can waste CPA on customers who won't or won't buy, and you need to know CLV to select customers to retain. When you have information about how much sales you will generate over the lifetime of your customers, you can decide how much money to invest in retaining them.

Explains predictive marketing specifically to increase CLV.

1) Predictive marketing is an activity that provides a more relevant and meaningful customer experience to customers at all touch points throughout the customer lifecycle, and improves customer loyalty and revenue. For recent tech startups such as Uber and Airbnb, predictive marketing is at the core of their business (Surge Pricing). In addition, Netflix's content recommendation and Amazon's product recommendation in the shopping cart are leading business cases using new technologies such as big data and AI.

2) Predictive marketing improves the precision of customer targeting and customer acquisition activities by identifying the channels that generate the most profitable customers, optimizing marketing spend, hyper-targeting specific micro-segments and increasing their conversion rates. You can design customer acquisition campaigns.

3) Predictive marketing can leverage personalized experiences to increase customer lifetime value. Through predictive marketing, you can improve the personalization, relevance, and timing of customer interactions, keep customers coming back, and maximize customer lifetime value.

4) Predictive marketing can increase understanding of customer retention and loyalty. You can improve the personalization, relevance, and timing of customer interactions. Make customers come back and maximize customer lifetime value.

5) Predictive marketing can increase understanding of customer retention and loyalty. Predicting when, why, and which customers will return or churn is a big challenge for many companies. Predictive marketing identifies and lists customers at risk of churn so you can take proactive steps to retain them.

6) Predictive marketing optimizes customer engagement. Predict who will respond to email promotions, what it will take to convert a casual customer into a buyer, and what discount will be needed to entice the customer to complete the transaction.

On the other hand, the traditional CLV prediction model (probabilistic model) predicts CLV from customer order transaction data. The three important inputs in the CLV model are recent purchases, frequency, and amount, which are called RFM (Recency, Frequency, and Monetary). Build a customer value prediction model based on when the customer last ordered, how often the customer buys, and how much the customer spends.

1) BTYD (Buy-Till-You-Die) model

In 1959, Ehrenberg presented in a thesis that the NBD model (Negative Binomial Distribution) is a consumer purchasing pattern, and presented an RFM-based probability model to predict the customer's future value based on the customer's order information in the past. The BTYD model learns the factor of the customer future value distribution to be most suitable for the RFM data set, and predicts the customer's future value by sampling from the distribution. What we want the model to predict is which customers are more likely to keep buying and not churn? (retention rate prediction), how many purchases will there be from each customer? (Predicting the number of purchases), and multiplying the number of purchases by the average purchase amount of the customer gives the customer value, and the CLV is calculated by multiplying this by the retention rate.

2) Pareto/NBD (PNBD) model

The Pareto distribution can be used to predict customer churn, and the NBD distribution can be used to predict future purchases. First, assuming that the customer lifespan model follows the Pareto distribution, general customers leave quickly after becoming customers, but certain customers become regular/loyal customers for a long time. Exp means exponential distribution or Pareto distribution, and μ ^c value is customer characteristic value and means customer churn rate.

Assuming that the number of purchases by a customer follows a Poisson distribution of k purchases during time t, the Poisson distribution is a kind of negative binomial distribution (NBD), where λ ^c denotes the number of purchases per unit time.

Customer life expectancy is predicted from the customer life span distribution, and the customer's expected number of purchases is predicted from the customer purchase distribution, and the customer lifetime value is obtained by multiplying the customer's expected number of purchases by the average purchase amount.

ML-based CLV prediction will be described. Traditional methods only use order transaction data, but machine learning (ML) models have the advantage of being able to reflect multiple characteristics other than RFM. Various information such as customer attributes, behavioral log data, and external environment can be used as predictive factors. Order and customer characteristics information before the base date are used for model training (Features), and orders after the base date are used to calculate target values (Labels). The label for the training data is CLV, for example, train a model to predict CLV for the past year with order information from the past two years, and predict CLV for next year with order information for the past year.

15 shows the overall characteristics of the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer life cycle according to the present invention. Referring to FIG. 15, the system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle according to the present invention includes customer attribute information including age and region information, customer behavior analysis data such as shopping carts, products of interest, and inquiry as well as purchases. Using , the customer value prediction model is learned by customer stage, but it is continuously learned through a time window that gradually increases the base date.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is only exemplary, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the attached claims.

110: customer information collection unit
120: life cycle classification unit
130: AUTOML part
140: CV calculation unit
150: model learning unit

Claims (10)

The customer information collection unit collects at least one of the customer's purchase history information as well as at least one of the customer's age and region including attribute information, store or site visits, shopping cart conversion rate, behavioral information including purchase history, and opinion information including comments. Collecting customer information including;
The life cycle classification unit calculates the sum of purchase amounts (CV) included in the customer purchase history information, and based on the calculated sum of purchase amounts and the number of purchases, the life cycle stage of the customer is new customers without purchases, first purchases, etc. Classifying into first-time purchase customers, general customers whose number of purchases is less than a first threshold, regular customers whose number of purchases is greater than the first threshold, and loyal customers whose customer lifetime value (CLV) is greater than or equal to a certain size;
The automatic machine learning (AUTOML) unit creates a learning model (M) that predicts future customer value with customer data for each of the new customers, first-time purchasers, regular customers, regular customers, and loyal customers, and the new customers and first-time purchasers , Learning by multiple learning cycles by regular customers, regular customers, and loyal customers (by customer life cycle), and learning through automatic machine learning (AUTOML) that automates coding, algorithm selection, and tuning tasks in the machine learning pipeline. doing;
obtaining a CV value realized by the customer life cycle by calculating a customer's actual purchase amount during the unit time after a predetermined unit time for each customer life cycle has elapsed, by a CV calculator; and
The model learning unit executes a plurality of learning models generated by the automatic machine learning (AUTOML) unit in a sliding time window method using the obtained CV value for each customer life cycle as the previous CV value, and the predicted CV value output is in advance If it is more than a set value, it includes a step of predicting a loyal customer,
When learning the learning model, a new prediction model (M) is created by taking over the previous learning model (Mprev) and learning it with a new learning data set (Data, CV), and the actual purchase amount by the customer is called CV , A method of automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle, selecting only customers whose difference from the predicted purchase amount, the predicted CV, is less than a preset threshold and learning the prediction model (M). delete delete The method of claim 1, wherein the learning model (M)
Characterized in that, by determining a specific point in time from customer information up to now, using data before the specific point in time as features, and using customer value (purchase amount) as a label from data after the specific point in time to predict the lifetime value of the customer, A method for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle. delete Customers that include at least one of the customer's purchase history information, at least attribute information including age and region of the customer, number of store or site visits, shopping cart conversion rate, behavioral information including purchase history, and opinion information including comments Customer information collection unit to collect information;
Calculate the sum of purchase amounts (CV) included in the customer purchase history information and classify the life cycle stage of the customer based on the calculated sum of purchase amounts and the number of purchases, a life cycle classification unit that classifies first-time purchase customers, general customers whose number of purchases is less than a first threshold, regular customers whose number of purchases is greater than the first threshold, and loyal customers whose lifetime value (CLV) is greater than or equal to a certain size;
Create learning models (M) that predict future customer value with customer data for each new customer, first-time purchase customer, regular customer, regular customer, and loyal customer (life cycle stage), and learn them through a plurality of learning cycles for each life cycle stage. and an AUTOML unit in which learning is performed through automatic machine learning (AUTOML) that automates coding, algorithm selection, and tuning in the machine learning pipeline;
After a predetermined unit time for each customer life cycle has elapsed, the customer calculates an actual purchase amount during the unit time to obtain a CV value realized for each customer life cycle. CV calculation unit; and
Predicting loyal customers through a plurality of learning models generated by the automatic machine learning (AUTOML) unit for each stage of the customer's life cycle using the obtained CV value for each customer life cycle as the previous CV value in a sliding time window method Including a model learning unit that does,
When learning the learning model, a new prediction model (M) is created by taking over the previous learning model (Mprev) and learning it with a new learning data set (Data, CV), and the actual purchase amount by the customer is called CV A system for automatically generating a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle, which selects only customers whose difference from the predicted purchase amount, the predicted CV, is less than a preset threshold and learns the prediction model (M). delete delete The method of claim 6, wherein the learning model (M)
Characterized in that, by determining a specific point in time from customer information up to now, using data before the specific point in time as features, and using customer value (purchase amount) as a label from data after the specific point in time to predict the lifetime value of the customer, A system that automatically generates a customer lifetime value (CLV) prediction model for each stage of a customer's lifecycle.

delete

Images