臺灣博碩士論文加值系統

近年分析模型與機械學習演算法的蓬勃發展，各式各樣演算法應用於不同領域，但零售業甚至電子商務零售業的數據分析應用卻不是非常普遍，價值創造也不如預期，接著又面對開源數據分析工具的發展，我認為應該能夠找出有效、也易於應用的演算法來建構回購預測模型，並解決近代電子商務常見的過度行銷問題，以及協助完成完整的CRM系統，因此產生了兩個研究目的：
(1)支持向量機與梯度增強決策樹兩種分類演算法，在跨產業的回購預測效果比較上，何者較優。
(2) 用機械學習方法(支持向量機與梯度增強決策樹)做回購預測，相較於過去常用的管理經驗法則建構的模型，是否能顯著提升回購預測的效果。
為了證明上述，本研究利用跨三個電商產業(美妝、女裝、食品)之實證研究，去建構以下三個模型(1)XGBOOST模型 (2)支持向量機模型 (3)管理經驗法模型，並比較每個模型預測出的F1score，去驗證機械學習的分類方法(支持向量機、梯度增強決策樹)應用在消費者回購行為的預測時，是不是能提供有效的預測結果。
實證結果證實以下發現：
支持向量機優於XGBOOST：由第五章第一節的結果可以明確表示，支持向量機在電商零售情境下的表現優於XGBOOST。
支持向量機方法整體占優：在不同產業、不同品牌情境下，支持向量機預測的效果皆優於管理經驗法模型的預測效果。
預測工具使用應適時適所：雖然在整理預測效果上(F1 score)，支持向量機有較優秀的表現，但細看至精確率(Precision Rate)與抓取率(Recall Rate)下，我們可以發現支持向量機有較好的精確率，而管理經驗模型的抓取率卻普遍優秀，所以管理者在使用工具時，應該要挑選適合搭配該檔行銷活動的工作座使用。
研究限制：整個實證研究過程中，發現了許多研究限制，如1.演算法極限 2.特徵變數有限 3.跨產業限制，大品牌小品牌差異 4.預測區間N天，這幾個限制都會影響實證研究成果，若在未來有辦法突破某些研究限制，相信預測結果會大大提升。

With the rising of data analytic model and Machine Learning algorithm, all kinds of algorithm have been applied on different area. Even so, the implication of data analysis in the e-commerce industry in Taiwan is still little, the value created by data analysis is not as much as expected. While the open source analytic tools being hot nowadays , there might be powerful and easy way to implement data analysis to build customer forecast model and solve the over marketing problems which is rising in modern e-commerce industry. Due to all above, there are two research problems:
(1)Which performs better on customer forecast in different industry? Support Vector Machine(SVM) or Gradient Boosting Decision Tree?
(2)Can Machine Learning techniques perform better than managerial heuristic method on customer forecast in different industry?
This research use transaction data from three different e-commerce and retail industry(cosmetics, women clothes, food product) to build three models as below. (1)XGBOOSTt model (2)Support Vector Machine Model (3) managerial heuristic model. This research will compare the performance of three different model base on F1 score。
The result as following：
Support Vector Machine performs better than XGBOOST.
Support Vector Machine performs better than managerial heuristic model in three different industry.
Different forecasting tools can be useful in different situation. Though base on F1 score, SVM always performs better, the Precision rate and Recall rate can tell different story. SVM always performs better on Precision rate while managerial heuristic method performs better on Recall rate. Marketers and managers should pick right tools base on different marketing situation.
Through the research, it found out several research restrictions that are algorithm limit, limited factor, industrial variation, scale impact and predict N days. These factors can cause severe impact on the model, yet the model efficiency can be improve by reducing these restrictions.

口試委員會審定書 #
誌謝 1
中文摘要 2
ABSTRACT 3
目錄 4
圖目錄 6
表目錄 7
第一章 緒論 8
第一節 研究背景與動機 8
1. 各類分析模型與機械學習演算法的發展 8
2. 台灣零售業數據分析應用程度與發展狀況 9
3. 開放平台分析工具盛行 10
4. 行銷手段多元，過度行銷問題產生 11
5. 回購預測模型，找出對的溝通時間點 11
第二節 研究目的與問題 12
第三節 研究流程與章節架構 12
第二章 文獻探討 14
第一節 RFM 模型 15
第二節 機械學習方法 16
第三章 資料描述與前處理 18
第一節 特徵變數說明與資料基本敘述統計 18
第二節 資料前處理 21
第四章 研究方法與設計 24
第一節 決策樹方法 25
1. 整體學習與Gradient Boosting Decision Tree 25
2. Extreme Gradient Boosting Tree(XGBOOST) 26
第二節 支援向量機學習法(Support Vector Machine) 28
第三節 管理經驗法 31
第四節 指標F1 score 32
第五章 實證研究 34
第一節 支持向量機 vs. 梯度增強決策樹 34
第二節 支持向量機 vs. 管理經驗法 38
第六章 結論與未來發展 40
參考文獻 42
附錄 44

Agresti, A. (2003). Categorical data analysis (Vol. 482). John Wiley & Sons.
Bartlett, M. S., Littlewort, G., Frank, M., Lainscsek, C., Fasel, I., & Movellan, J. (2005,June). Recognizing facial expression: machine learning and application tospontaneous behavior. In Computer Vision and Pattern Recognition, 2005. CVPR2005. IEEE Computer Society Conference on (Vol. 2, pp. 568-573). IEEE.
Belleflamme, P., Lambert, T., & Schwienbacher, A. (2014). Crowdfunding: Tapping theright crowd. Journal of business venturing, 29(5), 585-609.
Bolton, R. N. (1998). A dynamic model of the duration of the customer''s relationshipwith a continuous service provider: The role of satisfaction. Marketing science,17(1), 45-65.
Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992, July). A training algorithm foroptimal margin classifiers. In Proceedings of the fifth annual workshop onComputational learning theory (pp. 144-152). ACM.
Breiman, L. (1997). Arcing the edge. Technical Report 486, Statistics Department,University of California at Berkeley.
Breiman, L., Friedman, J. H., & Olshen, R. A. (85). stone, CJ (1984). Classification andregression trees.
Chen, T., & Guestrin, C. (2016, August). Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledgediscovery and data mining (pp. 785-794). ACM.
Fader, P. S., Hardie, B. G., & Lee, K. L. (2005). “Counting your customers” the easyway: An alternative to the Pareto/NBD model. Marketing science, 24(2), 275-284.
Fader, P. S., Hardie, B. G., & Lee, K. L. (2005). RFM and CLV: Using iso-value curvesfor customer base analysis. Journal of Marketing Research, 42(4), 415-430.
Friedman, J. H. (2001). Greedy function approximation: a gradient boosting machine.Annals of statistics, 1189-1232.
Galindo, J., & Tamayo, P. (2000). Credit risk assessment using statistical and machinelearning: basic methodology and risk modeling applications. ComputationalEconomics, 15(1), 107-143.
Huang, C. Y. (2012). To model, or not to model: Forecasting for customer prioritization.International Journal of Forecasting, 28(2), 497-506.
Kourou, K., Exarchos, T. P., Exarchos, K. P., Karamouzis, M. V., & Fotiadis, D. I.(2015). Machine learning applications in cancer prognosis and prediction.Computational and structural biotechnology journal, 13, 8-17.
Powers, D. M. (2011). Evaluation: from precision, recall and F-measure to ROC,informedness, markedness and correlation.
Samuel, Arthur L. (1959). "Some Studies in Machine Learning Using the Game ofCheckers". IBM Journal of Research and Development.
Taylor, J., King, R. D., Altmann, T., & Fiehn, O. (2002). Application of metabolomicsto plant genotype discrimination using statistics and machine learning.Bioinformatics, 18(suppl_2), S241-S248.
Vapnik, V. N., & Chervonenkis, A. Y. (1971). On the uniform convergence of relativefrequencies of events to their probabilities. Theory of Probability & ItsApplications, 16(2), 264-280.
Wübben, M., & Wangenheim, F. V. (2008). Instant customer base analysis: Managerialheuristics often “get it right”. Journal of Marketing, 72(3), 82-93.