TR-IIS-02-012    PDF format

Learning Hybrid Poisson Aspect Model For Personalized Recommendation

Advisor / Professor Chun-Nan Hsu, Student / Hao-Hsiang Chung


Abstract

Recommendations that really match customers' needs can boost sales. Researchers have proposed and evaluated many approaches for generating recommendations. In this thesis, we proposed a model-based collaborative approach, called Hybrid Poisson Aspect Model (HyPAM ). HyPAM is a hybrid system combining two probabilistic models, cluster and aspect, which model the relationship between customer clusters and product types. Given a new customer and his/her shopping record, HyPAM can estimate his/her degree of preference of each product item accurately. We use the EM algorithm to learn the parameters of HyPAM from customers' shopping records. To evaluate our approach, we apply HyPAM and two well-known recommender systems, GroupLens and IBM, to a shopping-record data set provided by a local supermarket. This data set contains 119,578 transactions of 32,266 distinguishable customers in a four-month period. We adopted two metrics, rank score and lift index, with four protocols, Given 2, Given 5, Given 10, and All but one. Under these evaluation metrics, experimental results show that HyPAM performs much better compared to the other two recommendation approaches for the given data set.

 

keywords: 個人化 personalized, 面向模型 aspect model, 混種式模型 hybrid model, 商品推薦 shopping recommendation, 資料探勘 data mining