登入選單

返回Google圖書搜尋

Hybrid and Dynamic Recommender Models for Consumer Decision Making and Marketing

Wei Deng

出版	University of Nebraska at Omaha, 2016
URL	http://books.google.com.hk/books?id=MoijzwEACAAJ&hl=&source=gbs_api

註釋

Recommender models are an information filtering system that provides users with customized suggestions based on their prior activity patterns. The early days of research in e-commerce recomendations primarily focused on static activity records, which restricts our understanding about customers' behaviors. Other data resources such as social media can reveal more dimensions about a customer. The other problem of static records is the absence of real-time information, which leads to the result of time-unawareness recommendation. Besides, the goal of many e-commerce recommendation approaches is to optimize accuracy. Accuracy only is inclined to cause the phenomenon of "filter bubbles". With these considerations, this dissertation presents a number of hybrid and dynamic algorithms aiming at better recommender systems, including moving recommender systems beyond prediction accuracy. Our research starts on how to incorporate graph information into existing Singular Value Decomposition algorithms. Three graph-based algorithms (adjust-SVD++, difference-SVD and topk-SVD) represents three ways to hybrid graph information in difference scenarios, We also propose a satisfaction maximization model, in which customer satisfaction is quantified, containing a utilitarian part and a hedonic part. Furthermore, the interaction between recommender algorithms and expert knowledge are discussed, so that recommender systems can be used to promote marketing sales. Two application scenarios are proposed: the first, a Trigger-Triggered model; the second a scenario of a churn prevention algorithm and related system design. The ultimate objective of algorithms in this dissertation is to achieve both improved marketing and the customer satisfaction. In particular, from the view of recommended contents, the goal of algorithms is to alleviate over-specification problem (recommended product is right, but frequency is intensive) and concentration problem (recommend hot product, lead to the risk of diversity diminish) while the accuracy is guaranteed. From the view of e-marketing, the goal of our research is to smooth the interaction between experts' knowledge and algorithms' output. From the view of the customer, the goal of design is to maximize their satisfaction and avoid "filter bubbles" problem.