| dc.contributor | 資管系 | |
| dc.creator (作者) | 周彥君 | |
| dc.date (日期) | 2021-09 | |
| dc.date.accessioned | 9-Jun-2026 13:15:16 (UTC+8) | - |
| dc.date.available | 9-Jun-2026 13:15:16 (UTC+8) | - |
| dc.date.issued (上傳時間) | 9-Jun-2026 13:15:16 (UTC+8) | - |
| dc.identifier.uri (URI) | https://ah.lib.nccu.edu.tw/item?item_id=182840 | - |
| dc.description.abstract (摘要) | 客戶回流率是一個在各產業都重要的問題，也因此一直是一個受到關注的研究議題。例如，在行銷領域透過機率假設建置的購買流失模型(buy till you die, BTYD)就是一個預測客戶回流的代表性的模型；這類模型我們統稱為低維度模型，因其僅需最近一次的消費時間(recency)以及消費次數(frequency)，即可透過機率模型的參數估計，而推斷出消費者的購買機率及流失機率，進而預測消費者下一期是否回流或是流失。相較於透過機率分配捕捉消費者行為的低維度模型，透過探索消費紀錄的機器學習模型則為高維度模型，其主要透過由消費紀錄所擷取多樣化的消費特徵(如：購買產品、金額、促銷金額及退貨等)，透過演算法探索特徵間複雜的關聯性，並以此做出客戶回流的預測。在近來人工智慧的浪潮下，透過機器學習來預測客戶回來獲得不少關注。本研究即透過一組超過50萬會員的線上零售資料，來探索以高維度資料的機器學習模型，以及透過機率模型假設消費者行為而僅需低維度輸入的BTYD模型，兩者間的比較同時進一步探討兩者間互補的可能性。在此情境下，我們使用Beta-Geometric/Beta-Bernoulli模型來代表線上零售的消費者行為，並將此模型所估計的參數作為高維度lasso regression的輸入特徵。此外，在高維度模型方面，我們亦將此簡單但可解釋的lasso regression，和更具有彈性但是黑箱的類神經網路做比較。透過此研究，對於客戶回流預測的文獻，我們提出兩大類型的模型的比較與整合；同時在實務上，整合的模型除了準確預測亦提供了決策所需的解釋性。 | |
| dc.description.abstract (摘要) | Customer retention has received broad research interests. In the field of marketing, the Buy till You Die (BTYD) models are perhaps the most representative techniques for customer retention prediction. Those models are low-dimensional in their inputs that typically involve only the recency and frequency of customer activities. Marketing researchers then employ various probability distributions to characterize customer lifetime, purchase intensity, and so forth, to model a repeated purchase process and a dropout process. Contrary to the low-dimensional probability modeling, a distinctly different class of predictive models for customer retention/churn is high-dimensional machine learning approaches. This class of models are inherent from the data mining avenue and high-dimensional in the sense that they rely on extracting a wide variety of features from customer transactions to make predictions. As the renaissance of machine learning lately, applying the method to the prediction of customer retention has gained traction. Using a large online retailing data composed of more than 500,000 members, we are inquired by how high-dimensional machine learning is compared to a completely different class of models that assume probabilities on the customer purchase process and requires low-dimensional inputs for parameter estimation. Furthermore, we explore the possibility of the two streams complement each other. Specifically, we use Beta-Geometric/Beta-Bernoulli model to represent low-dimensional probability model due to the discrete and non-contractual setting, and incorporate the parameter estimates of the model into the estimation of high-dimensional lasso regression. Such an interpretable lasso regression is further evaluated against artificial neural networks with flexible structures but low transparency of relationships of variables. The link of two steams of models would make contributions to the literature of predictive analytics for customer retention, and at the same time to practice as decision making requires both accurate and interpretable predictions. | |
| dc.format.extent | 116 bytes | - |
| dc.format.mimetype | text/html | - |
| dc.relation (關聯) | 科技部, MOST109-2410-H004-064, 109.08-110.07 | |
| dc.subject (關鍵詞) | 客戶回流; 預測性分析; BG/BB; Lasso Regression; 類神經網路; 機器學習 | |
| dc.subject (關鍵詞) | Customer Retention; Predictive Analytics; BG/BB; Lasso Regression; Artificial Neural Networks; Machine Learning | |
| dc.title (題名) | 客戶回流預測分析：高維度機器學習模型還是低維度消費行為機率模型？ | |
| dc.title (題名) | Predictive Analytics for Customer Retention: Low-Dimensional Probability Modeling or High-Dimensional Machine Learning? | |
| dc.type (資料類型) | report | |
