建立線上公司個人消費行為動態預警模式之研究

The Dynamic Predicting Model of Individual Consuming Behavior for Online Company

陳思翰(Suu-Han Chen)

指導教授 ： 羅淑娟

國立臺北科技大學/管理學院/工業工程與管理系所/碩士(2006年)

https://doi.org/10.6841/NTUT.2006.00179

摘要

本研究將建構一處理線上顧客狀態惡化的預警模式，此模式結合了指數累積和管制圖 (Exponential CUSUM Scheme) 與混合層級貝式模型 (Mixture Bayesian Hierarchical Model)，其中，混合層級貝式模型的主要目的為了區隔積極/非積極(Active/Inactive)消費行為並估計其對應之參數，對面對異質性的線上顧客，消費行為的監控有賴個別行為的標準，積極/非積極行為估計值將作為指數累積和管制圖之參考值 k (reference value) 的設計；累積和管制圖則可依以時間序列為基礎的次數軸進行檢視單一客戶是否存在狀態惡化趨向的預警。至於監控變數方面，在多位學者的實證分析下，認定間隔購買時間 (Interpurchase time) 可以充分地描繪消費行為的軌跡；也有學者提出，以近時作為最後一次購買 (Recency-of-Last-Purchase) 的準則亦可掌握消費者的近況，本研究將此兩重要的消費特徵點繪於最佳參數設計下的指數累積和管制圖。傳統的行為狀態判斷模式只是靜態地分析顧客的狀態，並且僅以數據式的報告或表格來呈現其結果；然而，累積和管制圖為我們帶來截然不同的預警架構，除了視覺化的圖形展現之外，隨著時間的流逝、新資訊的搜集，亦可動態地點繪新資訊於管制圖上，當某個指數累積和值 (Exponential CUSUM Score) 超出預定的管制界線時，便會產生失控的警訊，此的警訊的發生便提醒了管理者，某位顧客可能有狀態惡化的傾向，應及早做出防範對策。

關鍵字

顧客狀態 ； 顧客異質性 ； 有限混和模型 ； 層級貝式模型 ； 馬可夫鏈蒙地卡羅法 ； 指數累積和管制圖

並列摘要

In this thesis, we would like to establish a particular procedure for monitoring the online consuming behavior. This mechanism associates exponential CUSUM scheme with mixture Bayesian hierarchical model. Mixture Bayesian hierarchical model is used for getting objective parameters of individual active/inactive behaviors and then were applied at the optimal design of exponential CUSUM scheme. The empirical analysis of interpurchase time behavior of customers has received significant attention in recent years. Besides, a simple recency-of-last-purchase rule outperforms other complex models. Recency variable is contributive to judge whether customers still active or not. Our goal is to predict degenerate alarms by monitor the two important consuming characteristics, one is historical message and another is the latest new. As time goes by, we can depict the latest consuming behavior information on chart instantly. If there are some exponential CUSUM scores exceed the specific limit, out-of-control signals sound. It could announce a probable time point when customers’ states transfer to relative inactive and provides a friendly graphical version which can monitor customers’ states immediately and persistently. Dynamical monitor and graphical version characteristics are distinct from past consuming behavior predicting techniques which are statical analysis and numerical report only. A real-world case study from a website employ on our specific procedure that demonstrates the effectiveness of this specific mechanism. The valid experiment shows about 96% detective power while a customer’s behavior real transfer from active to inactive of this mechanism.

並列關鍵字

Customer State ； Customer Heterogeneity ； Finite Mixture Model ； Hierarchical Bayesian Model ； Markov Chain Monte Carlo Method ； Exponential Cumulative Sum Scheme

參考文獻

[1]. Allenby, G. M., Robert, P. L., and Jen, L. (1999), “A Dynamic Model of Purchase Timing with Application to Direct Marketing,” Journal of the American Statistical Association, Vol. 94, No. 446, pp. 365-373.

[2]. Blattberg, R. C., Getz, G., and Thomas, J. (2001), “Customer Equity: Building and Managing Relationships as Valuable Assets,” Boston, Mass Harvard Business School Press.

[5]. Gan, F. F. (1991), “An Optimal Design of CUSUM Quality Control Charts,” Journal of Quality Technology, Vol. 23, No. 4, pp. 279-286.

[7]. Gan, F. F. (1993), “An Optimal Design of CUSUM Quality Control Charts for Binomial Counts,” Journal of Applied Statistic, Vol. 20, pp. 445-460.

[8]. Gan, F. F. (1994a), “Computing Average Run Lengths for Exponential CUSUM Schemes,” Journal of Quality Technology, Vol. 26, No. 2, pp. 134-143.