摘要
現代社會的科技進步及資訊爆炸，導致資料庫行銷同時成為企業在做行銷時可資運用的利器，以及不得不使用的必備工具。其中，能將新舊產品推薦給顧客的推薦系統，更是企業能主動接觸顧客、維繫與顧客之間關係的強而有力手段。 而近年來，層級貝氏統計已被許多學者專家證實，能透過資料庫當中的資料，來準確預測顧客的相關行為或資訊。層級貝氏統計模式使用顧客的個人層次資訊來預測其後續資訊，並以全體顧客的總體層次資訊，來補足個人層次資訊的不足。透過上述的過程，層級貝氏統計模式得以預測顧客接下來的購買行為或價值遷移的過程。 本研究使用國內知名購物網站的資料庫資料，運用層級貝氏統計的普羅比模式，藉此建立顧客的個人專屬預測模式，並用來預測顧客之後可能會購買的商品。再以顧客的最後一筆保留紀錄，衡量此預測模式的準確性。 同時，本研究將會引入時間間隔的限制條件。因為前後兩次購物的時間間隔過長者，根據消費者行為的記憶提取相關理論而言，彼此之間並無法被認為是相關的。因此在處理資訊的過程中，將會據此淘汰不合理並篩選合理的顧客購買移轉記錄。此外，本研究會一同比較不同推薦方式的成效：條件機率的推薦方式及購物籃分析的推薦方式，並由此得出管理層面的意涵。 本研究的實證分析結論為，可先利用相關係數矩陣等方式，區分商品品項在分類上的粗細程度為何。若企業僅須推薦粗略的大品項推薦，則使用簡單的購物籃分析推薦即可；而企業若欲推薦細緻的個別商品推薦，則使用以層級貝氏統計為基礎的個人層次推薦，應能得到較佳的推薦成效果。
並列摘要
In modern society, the technology progression and information explosion, lead Database Marketing to become an edged and meanwhile essential tool, with which companies conduct marketing. Using Recommendation System, especially, is a powerful method for companies to vigorously contact customers, and to maintain customer relationships. Recently, Hierarchical Bayesian Statistic Models have been proven capable of predicting customers’ behavior or information precisely. HBSM utilize not only customers’ personal level of information to predict further relatives, but whole body’s overall level of information to supplement the insufficient of personal level. Through above processes, HBSM are able to predict customers’ further buying behaviors or value migration, etc. In this study, data in the database of a famous online shop in Taiwan is to be used. With the Probit model of HBSM, we can set up customers’ exclusively personal predict model, and predict customers’ next-buying items. And we will retain the customers’ actually final purchased items, in order to measure accuracy of the model. Moreover, the restriction of time interval will be introduced in this study. According to the memory retrieval theories, if the time interval between sequent buying behaviors is too long, then those behaviors could not be seen to be correlative. Therefore, above restriction will be used to sieve improper data out. Besides, the HB Probit methods will be compared about accuracy, with Conditional Probability method, and Market Base Analysis method. The conclusion in this study is that we can distinguish the degree of thickness of items’ separation by correlation matrix method. If the companies need only rough category recommendation, then they should use convenient methods like MBA method; and if the companies would like to conduct detailed item recommendation, then they should go through HB Probit method to attain outcome of personal exclusives.
參考文獻
7.張容華(2004)，「運用層級貝氏理論於顧客價值隨機模型之參數估計」，國立臺灣大學商學研究所未出版碩士論文。
9.許毓麟(2005)，「層級貝氏購物籃分析模型之研究」，國立臺灣大學國際企業學研究所未出版碩士論文。
3.Burke, R. (2002), “Hybrid Recommender Systems: Survey and Experiments”, pp. 332~335.
5.Cohen, A. L., Sanborn, A. N., & Shiffrin, R. M. (2008), “Model evaluation using grouped or individual data”. Psychonomic Bulletin & Review, 15, 692~712.
6.Copulsky, Jonathan R. & Michael J. Wolf. (1990), “Relationship Marketing: Positioning for the Future”, Journal of Business Strategy, pp.16~20.