| 摘要: | 現代零售業的數位軌跡已發展出商品、交易、行為及會員等多個維度，為了保持競爭力，企業採用以客戶為中心的策略，因此客戶關係管理變得越來越重要，在這種背景下，客戶流失的現象與原因值得被特別關注。馬可夫鏈 (Markov Chain) 屬於一種機率模型，用來表示一種狀態到另一種狀態的可能性，因此馬可夫鏈的應用適合描述客戶購買一種商品類別到另一種商品類別的可能性，提供了對未來結果的可能性估計。本研究提出了一種預測零售企業客戶流失的模型，其中使用客戶的購買商品序列與流失和非流失序列的相似性作為預測因子，使用馬可夫區別模型對購買事件的序列進行區分，馬可夫。個案分析中，本研究使用台灣零售企業提供的兩年資料，經過資料清洗與整理後得出四個變量，分別是客戶最近一次消費 (Recency)、客戶消費頻率 (Frequency)、客戶消費金額 (Monetary) 與客戶流失或非流失可能性 (Likelihood)，再來，使用邏輯迴歸分類技術，針對新客戶為研究對象，建立RFM與RFML客戶分類模型，最後，使用ROC曲線下面積 (Area Under Curve, AUC) 與準確性 (Accuracy) 評估分類模型辨別流失者的能力與正確分類的能力。研究結果顯示，馬可夫客戶流失預測模型具有足夠識別客戶流失的能力，在不同情境下，此模型皆具有穩定的準確率，在情境二中，企業常用的RFM分類模型，再加上Likelihood變數後，AUC與準確性皆有明顯的提升。本研究建立客戶流失預測模型，找出可能流失的客戶，並提供客戶關鍵特徵，幫助企業提前了解客戶的價值，以減少客戶流失的現象發生。;The digital trajectory of the modern retail industry has developed multiple dimensions such as products, transactions, behaviors, and memberships. To maintain competitiveness, companies adopt customer-centric strategies, making customer relationship management increasingly important. In this context, customer churn phenomena and their causes deserve special attention. Markov Chain is a probabilistic model used to represent the likelihood of transitioning from one state to another. Therefore, the application of Markov Chain is suitable for describing the probability of customers transitioning from purchasing one category of goods to another, providing estimates of future outcomes.
This study proposes a model for predicting customer churn in retail businesses. It utilizes the similarity between a customer′s purchase sequence and the churn and non-churn sequences as predictive factors. A Markov Discrimination Model is used to differentiate the sequence of purchase events. In a case study, this research uses two years of data provided by a Taiwanese retail company. After data cleaning and organization, four variables are obtained: Recency (the customer′s most recent purchase), Frequency (the customer′s purchase frequency), Monetary (the customer′s purchase amount) and Likelihood (the likelihood of churn or non-churn). Next, using logistic regression classification techniques, focusing on new customers as the research subjects, RFM and RFML customer classification models are established. Finally, the Receiver Operating Characteristic (ROC) curve′s Area Under the Curve (AUC) and accuracy are used to evaluate the models′ ability to identify churners and classify correctly.
The research results show that the Markov customer churn prediction model has sufficient ability to identify customer churn and maintains stable accuracy rates in different situations. In Scenario 2, when the commonly used RFM classification model is combined with the Likelihood variable, both AUC and accuracy show significant improvements. This study establishes a customer churn prediction model to identify potentially churned customers and provide key customer characteristics, helping businesses understand customer value in advance and reduce customer churn. |
