個人化新產品推薦系統：行為基礎與認知基礎模式之比較與整合

Abstract

產品推薦系統是落實消費者關係管理的一對 一行銷決策支援系統。在過去的研究中，產品推薦 系統大致可分成兩類，一是合作篩選系統 (collaborative filtering)，根據產品間之相關性進行 推薦；另是內容篩選系統(content filtering)，根據消 費者自我顯現(self-explicated)的偏好結構進行產品 推薦。這兩種推薦系統皆屬於總合層次，前者需要 大量購買紀錄方能得到較穩定的產品相關結構，後 者則需要集合相似的消費者資料方能得到穩定的 偏好結構，故皆無法充分反映消費者的異質性。此 外，合作篩選系統無法分析消費者選購產品的理由 或偏好結構，故無法進行新產品的推薦，只能就現 有產品進行關連性銷售(cross-selling)；傳統的內容 篩選系統無法探討個人偏好結構與人格特質的關 係，故無法進一步應用於對新客戶的產品推薦，只 能針對具有購買紀錄的舊消費者進行產品推薦。有 鑑於此，本研究擬以個人偏好結構(individual preference structure)為基礎，設計一套同時適用於 新產品與新消費者的產品推薦系統，期使消費者關 係管理之觀念能真正落實於企業之日常操作系統。 本研究根據消費者的主觀認知與客觀行為，分 別建立認知基礎與行為基礎的個人化偏好結構，試 圖比較二者是否具一致性，從而評估以主觀認知為 主的產品推薦系統的有效性。在主觀認知方面，為 降低受訪者的資訊處理負擔，本研究採取自我顯現 偏好，以問卷衡量之；在客觀行為方面，本研究以 相同受訪者的交易紀錄及虛擬產品構成該位消費 者的產品考慮集合，再以層級貝氏Probit 模式估計 個人化的偏好結構。實證結果顯示，認知基礎和行 為基礎的偏好結構不具一致性，這也帶出了以往利 用問卷方式去衡量購買態度，進而預測購買行為的 方式將會發生「適用性」及「預測效度」的問題。

Product recommendation system is a one-to-one marketing decision supporting system, which put customer relationship management into practice. In the past research, product recommendation systems fall into two classes. One is so-called collaborative filtering, which makes recommendation depending on correlation structure of all products. The other is known as content filtering, which makes recommendations on the basis of consumer’s self-explicated preference structure for product attributes. These two systems both are at aggregate level. The former needs dense purchases history data to get more stable correlation structure of products; the latter also needs pooling data sets of homogeneous customers to get more stable estimation of preference structure. Besides, for collaborative filtering provide few reasons for a recommendation and little information about preference structure of customers, it lacks the ability to make entirely new product recommendation but just make cross-selling among exiting products. The traditional content filtering does not analyze the relationship between individual preference structure and personality, so it cannot be applied to make recommendation for entirely new customers who provide no preference information. Therefore, this paper will design a product recommendation system suitable for both new items and new customers at individual level such that we can put the concept of customer relationship management into regular business operation systems. In marketing research field, conjoint analysis is a useful method to solve problems about individual preference structure. Most related research uses questionnaire-type survey data to explore customers' subjective recognition but not actual purchasing behavior. For the develop- ment and prevalence of database, it is easy to get purchase history data of every customer and purchase environment data such as sales promotion activity and advertisement on the point of purchase. According to concept of conjoint analysis, actual purchase behavior could be decomposed into preference structure and effects of purchase environment variables through proper models such as hierarchical Bayes probit model. Does behavior-based reference structure match with recognition-based? Which recommenda- tion system has better predictability? Does it depend on customer personality or purchase environment? All these topics will be discussed in this project.