透過您的圖書館登入
IP:3.89.163.120
  • 學位論文

顧客群預測之系統比較:行銷隨機模型vs機器學習

Systematic Comparisons of Customer Base Prediction: Marketing Stochastic Model vs. Machine Learning

指導教授 : 黃俊堯
本文將於2025/01/16開放下載。若您希望在開放下載時收到通知,可將文章加入收藏

摘要


顧客關係管理一直都是業界所關切的議題,其重點是以顧客為中心的關係管理策略。其核心目的在於保留高利潤顧客,維持更長的顧客關係,並讓已有顧客能奉獻更多交易流量。行銷科學家以行銷隨機模型為分析工具,資料分析師以機器學習為代表的分析工具。兩種迥異的研究方法皆能滿足顧客關係管理的需求,但過往的研究未曾對二者進行系統而全面的比較。因此,本文將關注於比較兩種研究方法在顧客群分析中預測能力的優勝狀況及其應用空間。 本文以Pareto/NBD (SMC)模型和Pareto/NBD (Abe)模型為行銷隨機模型的代表,以神經網絡、隨機森林、邏輯回歸和泊松回歸為機器學習的代表。本文採用遞迴比較策略對業界實際資料和模擬資料進行多個時點的實證分析。隨後,本文採用配對樣本"t" 檢定比較模型在不同比較階段的顯著性優勝。最後,透過決策樹分析得到合適的建模指引。依據資料集的數據豐富度,本文共進行基本行為變數情境與其他變數情境下的模型比較。 本文在基本行為變數情境下的比較中發現:Pareto/NBD (SMC)模型適合於中長觀測週期內的顧客留存與購買頻次預測;機器學習適合於短觀測週期內的顧客群預測。此外,機器學習相較於Pareto/NBD (SMC)在整體上具有更為的穩定的顧客群預測能力。綜合配對樣本"t" 檢定和決策樹分析的結果,當顧客人均可觀測生命週期長度較小時,機器學習在顧客留存與購買頻次中具有明顯的建模優勢。此外,若不考量顧客群的資料傾斜,則機器學習適合於短期內活躍的顧客群預測,Pareto/NBD (SMC)模型適合於相對活躍的客群的中長期和相對不活躍客群的中短期預測。若將資料傾斜納入考量,機器學習適合於短觀測週期的顧客群、中長期內不活躍的顧客群和長期內相對不活躍顧客群的預測。Pareto/NBD (SMC)模型適合於中長期內相對活躍的客群和中等觀測週期長度內不活躍的客群預測。 在其他變數情境的比較中,本文納入基本行為變數外的其他變數有助於改善行銷隨機模型在短期內的有偏預測,但會造成顧客購買頻次的高估。並且,行銷隨機模型無法借助其他變數顯著的提升預測能力。與之相對,機器學習的預測能力在其他變數情境下略有提升且在各個時點都表現穩定。然而,Pareto/NBD (SMC)模型的預測能力並未全部弱於其他變數下的機器學習算法。並且,Pareto/NBD (SMC)模型在中長期內的顧客群分析中具有較為明顯的優勢。 藉此,本文認為行銷隨機模型適合於基本行為變數情境下中長觀測週期的顧客群預測,機器學習適合於短期內的顧客群分析。在其他變數情境中,機器學習比行銷隨機模型具有更優的建模機會。

並列摘要


Customer relationship management has always been a topic of concern to the real business, and focused on customer-centric business strategies. The core purpose is to retain high-margin customers, maintain long-lasting customer relationships, and promote existing customers to conduct more transactions. Two different analytical perspectives offer different solutions: Marketing scientists harness the marketing stochastic model and data scientists utilize machine learning. Each solution can meet the needs of customer relationship management, but previous studies haven’t systematically and comprehensively compared these two methods under the context of customer base analysis. Therefore, this research aims to compare the superiority of these two methods in customer base prediction and define their application space. This research utilizes Pareto/NBD (SMC) model and Pareto/NBD (Abe) model as the representative of the marketing stochastic model, and uses neural network algorithm, random forest, logistic regression, and Poisson regression as the representative of machine learning. This paper adopts a recursive and expanding window strategy to conduct empirical analyses of industrial and simulated data in multiple comparison points. Subsequently, this paper divides the comparison points of each dataset into three stages and then uses the paired "t" test to verify the significant difference between models. Finally, the modeling guidelines are obtained through decision tree analysis. Based on the data richness, this research compares the models in the basic purchasing variable scenario and the additional variable scenario. In the basic purchasing variable scenario, this research finds the Pareto/NBD (SMC) model is more suitable for customer inactivity and purchase frequency predictions in the second- and third- comparison stages; machine learning has a better customer base prediction in the first-comparison stage. In addition, compared with Pareto/NBD (SMC), machine learning exists more stable predictive ability through comparison points. Based on the results of the paired t test and decision tree, when the average calibration length is short, machine learning has obvious modeling advantages in predictions of customer inactivity and purchase frequency. Under the Accuracy metric, machine learning is more suitable for the prediction of active customer base with small averaged recency; and the Pareto/NBD (SMC) model performs better in the relatively active customer base in the medium- and long- calibration periods and inactive customer base in the short- and medium- calibration periods. If the skewed classes problem is taken into consideration, machine learning is more suitable for the customer base in short calibration period, inactive customer base in medium- and long- calibration periods, and relatively inactive customer base in long calibration period. The Pareto/NBD (SMC) model has a predictive advantage in the relatively active customer base of the medium- and long- calibration periods and inactive customer base in the medium calibration period. In the additional variable scenario, the inclusion of additional variables besides the basic purchasing variables helps to improve the biased prediction of the Pareto/NBD (SMC) model in the short calibration period, but it causes a severe overestimation of purchase frequency. Moreover, the Pareto/NBD (Abe) model cannot significantly improve forecasting capabilities with the additional variables. In contrast, machine learning can improve the predictability and shows a more stable predictive performance in the additional variable scenario. However, machine learning with additional variables can’t exceed the Pareto/NBD (SMC) model in all comparison points. Furthermore, the Pareto/NBD (SMC) model can hold its predictive advantages over machine learning in the second- and third- comparison stages. Based on the empirical evidence, this research concludes that the marketing stochastic model is suitable for customer base prediction in the medium- and long- calibration periods with the basic purchasing variables, and machine learning is suitable for customer base analysis in the short calibration period. In the additional variable scenario, machine learning has better modeling capabilities than marketing stochastic models.

參考文獻


Abe, M. (2009), “Counting your customers” One by One: A Hierarchical Bayes Extension to the Pareto/NBD Model, Marketing Science, 28(3), 541-553.
Anand, G., A. H. Kazmi, P. Malhotra, L. Vig, P. Agarwal, and G. Shroff. (2015). Deep Temporal Features to Predict Repeat Buyers. Paper presented at the NIPS 2015 Workshop: Machine Learning for eCommerce.
Ang, L., and F. Buttle (2006), Customer Retention Management Processes: A Quantitative Study, European Journal of Marketing, 40(1/2), 83-99.
Ascarza, E., S. A. Neslin, O. Netzer, Z. Anderson, P. S. Fader, S. Gupta, . . . D. Neal (2018), In Pursuit of Enhanced Customer Retention Management: Review, Key Issues, and Future Directions, Customer Needs Solutions, 5(1-2), 65-81.
Baesens, B., S. Viaene, D. Van den Poel, J. Vanthienen, and G. Dedene (2002), Bayesian Neural Network Learning for Repeat Purchase Modelling in Direct Marketing, European Journal of Operational Research, 138(1), 191-211.

延伸閱讀