- 學位論文
基於多種資料群集之隱回饋跨領域推薦系統
Cross-domain recommender system based on multiple data integration with implicit feedback
摘要
集合多份推薦系統的資料群集有利於系統更瞭解使用者的喜好,系 統也因此可以推薦給使用者更好的商品清單,包括了跨站商品的推薦 以及原站的站內商品推薦。欲達成此目的,最簡單的方法就是直接將 多組資料視為一組資料,最後使用傳統的單領域推薦系統來學習使用 者喜好。然而,這可能會讓資料變得疏鬆以及多領域間的交互影響使 效用變低。大多數的跨領域系統是直接分析多組資料內使用者給的分 數,但是這種方式會因為同時去過兩個領域的重疊使用者太少而使得 結果變差。另外,這些方法都是使用人造切割成多領域的資料,並針 對使用者明回饋資料來分析。在這篇論文裡,我們提出了一個新的方 法,Content information as bridge across Multiple Domains Collaborative Filtering (CMDCF),是一個處理隱回饋多領域的推薦系統。在 CMDCF 裡,物品的資訊被用來連接多領域的一個依據而不只是使用者給的分 數。我們的實驗做在兩份真實的資料上面,與目前最好的方法來比較, 對於不同比例的重疊使用者,CMDCF 都能有效的猜出使用者喜好,並 且對於低重複使用者比例的資料有好的容忍度。除此之外,我們的方 法本身就可以減少在線上系統常會遇到的新使用者和新商品的問題。
並列摘要
Integrating data from multiple recommender systems helps us understand user preference more, which make us able to provider more useful recom- mendations including inter-domain and intra-domain recommendations. The simplest method is to merge multiple data set as one and directly adopt single domain recommendation methods. However, this will make the data more fragmentary and debase the overall performance. Most of previous related approaches directly analyze ratings from the integrated matrix to infer user performance, it always suffer from low percentage of overlapping users. Fur- thermore, they all demonstrate on synthetic data with explicit feedback. In this thesis, we propose a novel approach, Content information as bridge across Multiple Domains Collaborative Filtering (CMDCF), to effectively integrate multiple recommendation domains with implicit feedback. In CMDCF, con- tent information is utilized to construct links across multiple similar data sources. Experiments on two real online data sets with various amount of common overlapping users demonstrate the effectiveness and the high tolerance about overlapping user amount of CMDCF toward state-of-the-art methods. In ad- dition, our method inherently has good resistance against new user/item prob- lem, which usually occurs in the online environment.