TWM593025U

TWM593025U - Customer Behavior Prediction System

Info

Publication number: TWM593025U
Application number: TW108216776U
Authority: TW
Inventors: 王俊權; 黃逸琴; 周芳儀; 陳奕先; 魏得恩; 林得恩
Original assignee: 中國信託商業銀行股份有限公司
Priority date: 2019-12-17
Filing date: 2019-12-17
Publication date: 2020-04-01

Abstract

一種客戶行為預測系統，包含一儲存模組及一電連接該儲存模組的處理模組；該儲存模組儲存有多筆相關於一待預測客戶於先前做出之多種不同行為且各自對應有一先前時間點的先前行為資料，該處理模組根據該等先前行為資料，獲得一包含多個對應該等先前行為資料之先前參數的先前序列，該等先前參數依照該等先前時間點排序，根據該先前序列，利用一用於預測客戶未來行為的客戶行為預測模型，獲得一包含多個指示出該待預測客戶未來可能各種行為之預測參數的預測序列。A customer behavior prediction system includes a storage module and a processing module electrically connected to the storage module; the storage module stores a plurality of different behaviors related to a customer to be predicted in the past and each corresponds to one According to the previous behavior data of the previous time point, the processing module obtains a previous sequence including a plurality of previous parameters corresponding to the previous behavior data according to the previous behavior data, the previous parameters are sorted according to the previous time points, according to The previous sequence uses a customer behavior prediction model for predicting the customer's future behavior to obtain a prediction sequence that includes a plurality of prediction parameters indicating possible future behaviors of the customer to be predicted.

Description

客戶行為預測系統Customer Behavior Prediction System

本新型是有關於一種行為預測系統，特別是指一種利用機器學習的行為預測系統。The present invention relates to a behavior prediction system, especially a behavior prediction system using machine learning.

在銀行業中現有的行銷方式以產品經理人經驗為主，以人工檢索客戶過去消費行為資料並挑選出目標客群，並透過後續的行銷活動對所挑選出目標客群進行驗證，以判定所挑選目標客群的優劣，而產品經理人最後再依據挑選目標客群之優劣的反饋，提升自身的挑選經驗。The existing marketing methods in the banking industry are based on the experience of product managers, manually retrieving customer past consumer behavior data and selecting target customer groups, and verifying the selected target customer groups through subsequent marketing activities to determine Select the pros and cons of the target customer group, and the product manager finally improves his selection experience based on the feedback of the pros and cons of the target customer group.

然而，此種人工挑選目標客群方式不但沒有一個固定的標準，在培訓產品經理人方面也耗費過多時間成本，所挑選出來目標客群精確度也不足以達成金融商品推銷的目的。However, this manual selection of target customer groups not only does not have a fixed standard, but also takes too much time and cost in training product managers, and the accuracy of the selected target customer groups is not sufficient to achieve the purpose of financial commodity marketing.

有鑑於此，勢必須提出一種全新解決方案，以提升挑選出來目標客群的精確度。In view of this, it is imperative to propose a completely new solution to improve the accuracy of selecting target customer groups.

因此，本新型之目的，即在提供一種基於機器學習以提升挑選出來目標客群的精確度的客戶行為預測系統。Therefore, the purpose of the new model is to provide a customer behavior prediction system based on machine learning to improve the accuracy of the selected target customer group.

於是，本新型客戶行為預測系統包含一儲存模組及一電連接該儲存模組的處理模組。Therefore, the new customer behavior prediction system includes a storage module and a processing module electrically connected to the storage module.

該儲存模組儲存有多筆相關於一待預測客戶於先前做出之多種不同行為且各自對應有一先前時間點的先前行為資料。The storage module stores a plurality of pieces of previous behavior data related to a variety of different behaviors of a customer to be predicted in the past and each corresponding to a previous time point.

該處理模組根據該等先前行為資料，獲得一包含多個對應該等先前行為資料之先前參數的先前序列，該等先前參數依照該等先前時間點排序，根據該先前序列，利用一用於預測客戶未來行為的客戶行為預測模型，獲得一包含多個指示出該待預測客戶未來可能各種行為之預測參數的預測序列。The processing module obtains a previous sequence including a plurality of previous parameters corresponding to the previous behavior data according to the previous behavior data, the previous parameters are sorted according to the previous time points, and according to the previous sequence, a The customer behavior prediction model for predicting the future behavior of the customer obtains a prediction sequence containing a plurality of prediction parameters indicating various possible future behaviors of the customer to be predicted.

本新型之功效在於：藉由該待預測客戶之先前行為所對應的該先前序列，利用用於預測客戶未來行為的該客戶行為預測模型，以獲得該未來預測序列，而該未來預測序列所包含的每一預測參數皆指示出該待預測客戶未來可能之行為，而銀行便能根據該待預測客戶未來可能之行為判定該待預測客戶是否為目標客戶，大大地提升了挑選出目標客群精確度。The effect of the present invention lies in that: by using the previous sequence corresponding to the previous behavior of the customer to be predicted, the customer behavior prediction model used to predict the customer's future behavior is used to obtain the future prediction sequence, and the future prediction sequence includes Each forecast parameter indicates the possible future behavior of the customer to be predicted, and the bank can determine whether the customer to be predicted is the target customer according to the future behavior of the customer to be predicted, greatly improving the accuracy of selecting the target customer group degree.

參閱圖1，本新型客戶行為預測系統1之一實施例經由一通訊網路200連接一客戶端2。Referring to FIG. 1, an embodiment of the new customer behavior prediction system 1 is connected to a client 2 via a communication network 200.

該客戶行為預測系統1包含一連接該通訊網路200的系統端通訊模組11、一系統端儲存模組12，以及一電連接該系統端通訊模組11與該系統端儲存模組12的系統端處理模組13。The customer behavior prediction system 1 includes a system-side communication module 11 connected to the communication network 200, a system-side storage module 12, and a system electrically connecting the system-side communication module 11 and the system-side storage module 12 End-processing module 13.

該系統端儲存模組12儲存有多筆對應多個客戶的訓練資料，每一訓練資料包含所對應之客戶於先前做出該等不同行為的多筆訓練先前行為資料及多筆訓練預測行為資料，以及多個對應各種目標客戶行為之可推薦參數及其對應的一推薦資訊。其中，每一筆訓練先前行為資料皆對應有一訓練先前時間點，每一筆訓練預測行為資料皆對應有一訓練預測時間點，同一筆訓練資料之任一訓練預測時間點皆晚於該等訓練先前時間點。其中，每一推薦資訊係為金融商品資料，但不以此為限。The system-side storage module 12 stores a plurality of pieces of training data corresponding to a plurality of customers, and each training data includes a plurality of pieces of previous training behavior data and a plurality of pieces of predicted training behavior data that the corresponding customer previously made these different behaviors , And multiple recommendable parameters corresponding to various target customer behaviors and a corresponding recommendation information. Wherein, each training prior behavior data corresponds to a previous training time point, each training prediction behavior data corresponds to a training prediction time point, and any training prediction time point of the same training data is later than these training previous time points . Among them, each recommendation information is financial commodity information, but not limited to this.

該客戶端2包含一連接該通訊網路200的客戶端通訊模組21、一客戶端顯示模組22，以及一電連接該客戶端通訊模組21與該客戶端顯示模組22的客戶端處理模組23。The client 2 includes a client communication module 21 connected to the communication network 200, a client display module 22, and a client process electrically connecting the client communication module 21 and the client display module 22 Module 23.

在該實施例中，該客戶行為預測系統1之實施態樣例如為一個人電腦、一伺服器或一雲端主機，但不以此為限。In this embodiment, the implementation of the customer behavior prediction system 1 is, for example, a personal computer, a server, or a cloud host, but not limited to this.

在該實施例中，該客戶端2之實施態樣例如為一個人電腦、一智慧型手機或一平板電腦，但不以此為限。In this embodiment, the implementation of the client 2 is, for example, a personal computer, a smart phone, or a tablet computer, but not limited to this.

以下將藉由本新型客戶行為預測系統1執行一客戶行為預測方法來說明該系統端通訊模組11、該系統端儲存模組12、該系統端處理模組13，以及該客戶端2各元件的運作細節。該客戶行為預測方法包含一模型訓練程序，以及一預測推薦程序。The following will describe the system-side communication module 11, the system-side storage module 12, the system-side processing module 13, and the components of the client 2 by using the new customer behavior prediction system 1 to execute a customer behavior prediction method Operation details. The customer behavior prediction method includes a model training program and a prediction recommendation program.

參閱圖2，該模型訓練程序包含步驟51~步驟53。Referring to FIG. 2, the model training procedure includes steps 51 to 53.

在步驟51中，對於每一訓練資料，該系統端處理模組13根據該訓練資料中的該等訓練先前行為資料，獲得一包含多個對應該等訓練先前行為資料之訓練先前參數的訓練先前序列。其中，該等訓練先前參數依照該等訓練先前時間點排序。In step 51, for each training data, the system-side processing module 13 obtains a training previous containing a plurality of training previous parameters corresponding to the training previous behavior data according to the training previous behavior data in the training data sequence. Wherein, the previous training parameters are sorted according to the previous training time points.

在步驟52中，對於每一訓練資料，該系統端處理模組13根據該訓練資料中的該等訓練預測行為資料，獲得一包含多個對應該等訓練預測行為資料之訓練預測參數的訓練預測序列，該等訓練預測參數依照該等訓練預測時間點排序。In step 52, for each training data, the system-side processing module 13 obtains a training prediction including a plurality of training prediction parameters corresponding to the training prediction behavior data according to the training prediction behavior data in the training data Sequence, the training prediction parameters are ordered according to the training prediction time points.

值得特別說明的是，對於每一訓練資料，當該等訓練先前行為資料及該等訓練預測行為資料皆係為客戶***消費資訊時，則該等訓練先前參數（訓練資料之輸入）及該等訓練預測參數（訓練資料之輸出）皆係為商戶類別代碼（MCC CODE，Merchant Category Code，用於標明商家提供的商品或服務的類型）時；而，當該等訓練先前行為資料及該等訓練預測行為資料皆係為客戶網路瀏覽資訊時，則該等訓練先前參數（訓練資料之輸入）及該等訓練預測參數（訓練資料之輸出）皆係為網站代碼，但不以上述舉例為限。其中，每一網路瀏覽資訊中的網址皆可對應於一網站代碼。It is worth noting that for each training data, when the training previous behavior data and the training prediction behavior data are customer credit card consumption information, then the training previous parameters (input of training data) and these The training prediction parameters (the output of training data) are all merchant category codes (MCC CODE, Merchant Category Code, used to indicate the type of goods or services provided by the merchant); and, when such training prior behavior data and such training When the predicted behavior data are all the information browsed by the customer on the Internet, the training previous parameters (input of training data) and the training prediction parameters (output of training data) are all website codes, but not limited to the above examples . Wherein, the web address in each web browsing information can correspond to a web site code.

在步驟53中，該系統端處理模組13根據該等訓練資料，利用循環神經網路，獲得一用於預測客戶未來行為的客戶行為預測模型。其中，循環神經網路係為遞歸神經網路（RNN，Recurrent Neural Networks）或長短期記憶模型（LSTM，Long Short-Term Memory），但不以此為限。In step 53, the system-side processing module 13 uses a recurrent neural network based on the training data to obtain a customer behavior prediction model for predicting the customer's future behavior. Among them, the recurrent neural network is a recurrent neural network (RNN, Recurrent Neural Networks) or a long short-term memory model (LSTM, Long Short-Term Memory), but not limited to this.

參閱圖3，該預測推薦程序包含步驟61~步驟65。Referring to FIG. 3, the prediction recommendation procedure includes steps 61 to 65.

在步驟61中，該系統端處理模組13根據多筆相關於一待預測客戶於先前做出之多種不同行為且各自對應有一先前時間點的先前行為資料，獲得一包含多個對應該等先前行為資料之先前參數的先前序列。其中，該等先前參數依照該等先前時間點排序。In step 61, the system-side processing module 13 obtains a plurality of previous behavior data related to a plurality of different behaviors of a customer to be predicted in the past and each corresponding to a previous time point. The previous sequence of the previous parameters of the behavioral data. Among them, the previous parameters are sorted according to the previous time points.

在步驟62中，該系統端處理模組13根據該先前序列，利用所訓練出之該客戶行為預測模型，獲得一包含多個指示出該待預測客戶未來可能各種行為之預測參數的預測序列。In step 62, the system-side processing module 13 uses the trained customer behavior prediction model according to the previous sequence to obtain a prediction sequence including a plurality of prediction parameters indicating possible future behaviors of the customer to be predicted.

值得特別說明的是，當該等先前行為資料係為客戶***消費資訊，且該等先前參數係為商戶類別代碼時，則該等預測參數係為商戶類別代碼，此時該等預測參數即表示未來客戶可能會消費的商店類別；而，當該等先前行為資料係為客戶網路瀏覽資訊，且該等先前參數係為網站代碼時，則該等預測參數係為網站代碼，此時該等預測參數即表示未來客戶可能會點選瀏覽的網站，但不以上述舉例為限。其中，每一網路瀏覽資訊中的網址皆可對應於一網站代碼。It is worth noting that when the previous behavior data is customer credit card consumption information, and the previous parameters are merchant category codes, then the prediction parameters are merchant category codes, and then the prediction parameters represent The type of store that the customer may consume in the future; and when the previous behavior data is the customer's Internet browsing information and the previous parameters are the website code, then the prediction parameters are the website code The prediction parameters mean that in the future, the customer may click on the website to browse, but not limited to the above example. Wherein, the web address in each web browsing information can correspond to a web site code.

在步驟63中，該系統端處理模組13判定該預測序列的該等預測參數中是否存在與該等可推薦參數中之任一者相同的至少一待推薦參數。當該系統端處理模組13判定出不存在任一待推薦參數時，結束該預測推薦程序；當該系統端處理模組13判定出存在該至少一待推薦參數時，進行流程步驟64。In step 63, the system-side processing module 13 determines whether there is at least one parameter to be recommended that is the same as any of the recommendable parameters among the prediction parameters of the prediction sequence. When the system-side processing module 13 determines that there is no parameter to be recommended, the prediction recommendation process is ended; when the system-side processing module 13 determines that the at least one parameter to be recommended exists, process step 64 is performed.

在步驟64中，該系統端處理模組13透過該系統端通訊模組11將每一待推薦參數所對應之推薦資訊傳送至該客戶端2。In step 64, the system-side processing module 13 transmits the recommendation information corresponding to each parameter to be recommended to the client 2 through the system-side communication module 11.

在步驟65中，該客戶端處理模組23在透過該客戶端通訊模組21接收到每一待推薦參數所對應之推薦資訊後，將每一待推薦參數所對應之推薦資訊顯示於該客戶端顯示模組22。In step 65, after receiving the recommendation information corresponding to each parameter to be recommended through the client communication module 21, the client processing module 23 displays the recommendation information corresponding to each parameter to be recommended to the client端Displaymodule 22.

綜上所述，本新型客戶行為預測系統1，藉由該客戶端處理模組23對預先獲得的每一訓練資料進行前處理，以獲得對應該訓練資料且包含該等訓練先前參數的該訓練先前序列，及對應該訓練資料且包含該等訓練預測參數的該訓練預測序列，並利用循環神經網路，將該訓練先前序列作為輸入，而將該訓練預測序列作為預期輸出，訓練出該客戶行為預測模型；接著，該客戶端處理模組23根據該待預測客戶之先前行為所對應的該先前序列，利用所訓練出的該客戶行為預測模型，獲得該未來預測序列，而該未來預測序列所包含的每一預測參數皆指示出該待預測客戶未來可能之行為，並根據該待預測客戶未來可能之行為推薦相對應的金融商品，而銀行便能根據該待預測客戶未來可能之行為判定該待預測客戶是否為目標客戶，進而大大地提升了挑選出目標客群精確度，使得所推薦的金融商品更能符合該待預測客戶的需求。因此，確實能達成本新型之目的。In summary, in the new customer behavior prediction system 1, the client processing module 23 performs pre-processing on each training data obtained in advance to obtain the training corresponding to the training data and including the previous training parameters The previous sequence, and the training prediction sequence corresponding to the training data and containing the training prediction parameters, and using a recurrent neural network, using the training previous sequence as input and the training prediction sequence as the expected output to train the client Behavior prediction model; then, the client processing module 23 obtains the future prediction sequence using the trained customer behavior prediction model according to the previous sequence corresponding to the previous behavior of the customer to be predicted, and the future prediction sequence Each prediction parameter included indicates the possible future behavior of the customer to be predicted, and recommends corresponding financial products based on the future behavior of the customer to be predicted, and the bank can make a judgment based on the future behavior of the customer to be predicted Whether the customer to be predicted is a target customer, thereby greatly improving the accuracy of selecting the target customer group, so that the recommended financial products can better meet the needs of the customer to be predicted. Therefore, it can indeed achieve the purpose of new cost.

惟以上所述者，僅為本新型之實施例而已，當不能以此限定本新型實施之範圍，凡是依本新型申請專利範圍及專利說明書內容所作之簡單的等效變化與修飾，皆仍屬本新型專利涵蓋之範圍內。However, the above are only examples of the new model. When the scope of the new model cannot be limited by this, any simple equivalent changes and modifications made according to the patent application scope and patent specification content of the new model are still regarded as Within the scope of this new patent.

200:通訊網路 1:客戶行為預測系統 11:系統端通訊模組 12:系統端儲存模組 13:系統端處理模組 2:客戶端 21:客戶端通訊模組 22:客戶端顯示模組 23:客戶端處理模組 51~53:步驟 61~65:步驟200: communication network 1: Customer behavior prediction system 11: System-side communication module 12: System-side storage module 13: System-side processing module 2: client 21: Client communication module 22: Client display module 23: Client processing module 51~53: steps 61~65: steps

本新型之其他的特徵及功效，將於參照圖式的實施方式中清楚地呈現，其中：圖1是一方塊圖，說明本新型客戶行為預測系統的一實施例；圖2是一流程圖，說明該實施例所執行之一客戶行為預測方法的一模型訓練程序；及圖3是一流程圖，說明該客戶行為預測方法的一預測推薦程序。 Other features and functions of the present invention will be clearly presented in the embodiments with reference to the drawings, in which: FIG. 1 is a block diagram illustrating an embodiment of the new customer behavior prediction system; FIG. 2 is a flowchart illustrating a model training procedure of a customer behavior prediction method executed by the embodiment; and FIG. 3 is a flowchart illustrating a prediction recommendation procedure of the customer behavior prediction method.

200:通訊網路 200: communication network

1:客戶行為預測系統 1: Customer behavior prediction system

11:系統端通訊模組 11: System-side communication module

12:系統端儲存模組 12: System-side storage module

13:系統端處理模組 13: System-side processing module

2:客戶端 2: client

21:客戶端通訊模組 21: Client communication module

22:客戶端顯示模組 22: Client display module

23:客戶端處理模組 23: Client processing module

Claims

一種客戶行為預測系統，包含：一儲存模組，儲存有多筆相關於一待預測客戶於先前做出之多種不同行為且各自對應有一先前時間點的先前行為資料；及一處理模組，電連接該儲存模組，根據該等先前行為資料，獲得一包含多個對應該等先前行為資料之先前參數的先前序列，該等先前參數依照該等先前時間點排序，根據該先前序列，利用一用於預測客戶未來行為的客戶行為預測模型，獲得一包含多個指示出該待預測客戶未來可能各種行為之預測參數的預測序列。 A customer behavior prediction system, including: A storage module that stores multiple pieces of previous behavior data related to a variety of different behaviors of a customer to be predicted in the past and each corresponding to a previous time point; and A processing module, electrically connected to the storage module, and according to the previous behavior data, obtain a previous sequence including a plurality of previous parameters corresponding to the previous behavior data, the previous parameters are sorted according to the previous time points, according to The previous sequence uses a customer behavior prediction model for predicting the customer's future behavior to obtain a prediction sequence that includes a plurality of prediction parameters indicating possible future behaviors of the customer to be predicted.

如請求項1所述的客戶行為預測系統，還包含一經由一通訊網路連接至一客戶端且電連接該處理模組的通訊模組，該儲存模組還儲存有多個對應各種目標客戶行為之可推薦參數及其對應的一推薦資訊，該處理模組判定該預測序列的該等預測參數中是否存在與該等可推薦參數中之任一者相同的至少一待推薦參數，當該處理模組判定出存在該至少一待推薦參數時，該處理模組透過該通訊模組將每一待推薦參數所對應之推薦資訊傳送至該客戶端。The customer behavior prediction system according to claim 1, further comprising a communication module connected to a client via a communication network and electrically connected to the processing module, the storage module further stores a plurality of corresponding customer behaviors The recommendable parameter and its corresponding recommendation information, the processing module determines whether there is at least one parameter to be recommended that is the same as any of the recommendable parameters among the predictive parameters of the prediction sequence, when the processing When the module determines that the at least one parameter to be recommended exists, the processing module transmits the recommendation information corresponding to each parameter to be recommended to the client through the communication module.

如請求項1所述的客戶行為預測系統，其中，該儲存模組還儲存有多筆對應多個客戶的訓練資料，每一訓練資料包含所對應之客戶於先前做出該等不同行為的多筆訓練先前行為資料及多筆訓練預測行為資料，每一筆訓練先前行為資料皆對應有一訓練先前時間點，每一筆訓練預測行為資料皆對應有一訓練預測時間點，同一筆訓練資料之任一訓練預測時間點皆晚於該等訓練先前時間點，對於每一訓練資料，該處理模組根據該訓練資料中的該等訓練先前行為資料，獲得一包含多個對應該等訓練先前行為資料之訓練先前參數的訓練先前序列，該等訓練先前參數依照該等訓練先前時間點排序，對於每一訓練資料，該處理模組根據該訓練資料中的該等訓練預測行為資料，獲得一包含多個對應該等訓練預測行為資料之訓練預測參數的訓練預測序列，該等訓練預測參數依照該等訓練預測時間點排序，該處理模組根據該等訓練資料，利用循環神經網路，獲得該客戶行為預測模型。The customer behavior prediction system according to claim 1, wherein the storage module further stores a plurality of pieces of training data corresponding to a plurality of customers, and each training data includes the number of times the corresponding customer has previously performed these different behaviors Pen training previous behavior data and multiple training prediction behavior data, each training previous behavior data corresponds to a previous training time point, each training prediction behavior data corresponds to a training prediction time point, any training prediction of the same training data The time points are later than the previous training time points. For each training data, the processing module obtains a training previous containing multiple corresponding training previous behavior data according to the training previous behavior data in the training data The training previous sequence of the parameters, the training previous parameters are sorted according to the training previous time points, and for each training data, the processing module obtains a plurality of correspondences according to the training prediction behavior data in the training data The training prediction sequence of training prediction parameters of the training prediction behavior data, the training prediction parameters are sorted according to the training prediction time points, the processing module obtains the customer behavior prediction model based on the training data, using a recurrent neural network .

如請求項1所述的客戶行為預測系統，其中，該等先前行為資料係為客戶***消費資訊，該等先前參數係為商戶類別代碼，該等預測參數係為商戶類別代碼。The customer behavior prediction system according to claim 1, wherein the previous behavior data is customer credit card consumption information, the previous parameters are merchant category codes, and the prediction parameters are merchant category codes.

如請求項1所述的客戶行為預測系統，其中，該等先前行為資料係為客戶網路瀏覽資訊，該等先前參數係為網站代碼，該等預測參數係為網站代碼。The customer behavior prediction system according to claim 1, wherein the previous behavior data is customer web browsing information, the previous parameters are website codes, and the prediction parameters are website codes.