線上新產品推薦系統－以亞馬遜網路書店為例

Abstract

本研究是以資料庫行銷的技術來剖析每一位顧客的真實消費狀況，利用顧客過去的交易紀錄，來推測其未來的購買行為，同時導入顧客推薦系統的概念，將顧客視為有不同偏好的獨立個體，分別推薦顧客不同的產品，以期能為企業贏得最大的利潤。 全文通篇將資料庫內所有的顧客劃分為兩群組：第一群顧客是交易紀錄超過一筆，顯示為曾經重複購買者，被視為「舊顧客」；第二群顧客是僅有一筆紀錄，則視其為「新顧客」。從舊顧客的資料中個別挑選出最近期的一次紀錄，作為樣本內估計顧客購買偏好，及分析顧客購買行為的依據；剩餘的所有資料則分別用以預測舊顧客與新顧客的購買機率。一但了解顧客的購買偏好與發生交易的機率後，便可以對顧客進行最直接的產品推薦。 首先由顧客最初的購買行為之需求端切入，將顧客的購買偏好依照相異程度的不同，導入統計方法來分析，並且估計顧客購買偏好，來預測顧客下一次的購買；若預測顧客應該會購買，而資料顯示顧客確實有購買紀錄時，則謂之「擊中」。而後便比較此三種不同的新產品推薦模式，以累積擊中率來判斷其模式的優劣，分別包含平均機率法推薦模式、總合邏吉斯推薦模式，以及在行銷理論模型中，充分展現預測效果最佳的層級貝氏邏吉斯推薦模式。 最終實證後所得出的結論顯示，層級貝氏邏吉斯推薦模式的表現的確最好，但是仍有些許外在因素干擾其預測結果，倘若加以改善則可擁有更好的預測能力。本研究逐一針對各推薦模式的實證結果給予正反向評判，期望能由顧客關係管理中精準的資料庫行銷技術當作指引，一方面提供顧客客製化的推薦服務等，另一方面則帶領企業獲得顧客的終身價值，協助企業帶來更大的商機，創造買賣雙贏的局面。

It’s getting obvious that customer relationship management could be viewed as a lethal weapon. To give the customers exactly what they want in affordable price can easily enhance the customer satisfaction. And this study is based on database marketing techniques, which is the key point of CRM, to analyze each customer’s purchasing behavior. By examining the transaction records, we’ll predict each customer’s next purchasing behavior, and apply the concept of customer recommendation systems to customize their recommended products. We believe that with one strong and precise recommendation system, we could encourage cross-buying, develop customer loyalty, and finally improve the customer retention, which would lead to great profits. In this research, we hold two purposes, one is to find out the online books buying preference, the other is to compare the different kinds of recommendation systems. We separate all customers into two groups by their repetitive purchasing in turn representing“Old Customers” and “New Customers”. Later we try three types of statistical models, “the Common Average Method”, “the Aggregate Logit Recommendation System”, and the “Hierarchical Bayesian Logit Recommendation System”, and see which one of them can perform the best in the accumulated hit ratio for predicting customers purchasing possibilities. The study result shows several buying habits on different cluster of customers, for example, people with only high school educational backgrounds prefer buying “Business and Finance” to “Lifestyle” types of books, they would like to buy books that are thicker, and so on. In addition to the purchasing behavior, we also found that Hierarchical Bayesian Logit Recommendation System do the best prediction, just like the Hierarchical Bayes theory proposed, no matter for old customers or new customers. It’s is quite evident that more and more customers now they share heterogeneity needs,. In that, to best serve all individual’s need, the company better keep on customizing the recommendation system.