個人化書籍推薦方法之推薦品質比較

Abstract

隨著網際網路的興起，使用者大量依賴網路，我們在各個網路平台上分享張貼訊息，同時也透過入口網站找尋資料，但在這樣資訊過多的網路世界中，使用者難以從中找到對的資訊，推薦系統因而成為改善此問題的重要方法。目前有越來越多的針對推薦系統的學術研究，為了改善推薦精確度，學者們從不同面向來探討推薦系統的影響。 本研究資料集取自全球最大的書籍社群網站Goodreads ，我們提出了一個新的混合推薦方法，此方法結合了傳統使用者協同過濾及內容推薦兩種方法。在協同過濾方法中，我們加入三種不同的影響力分析，分別為評論影響力、朋友影響力及跟隨者影響力；在內容推薦方法中，我們使用潛在狄利克里分配分析使用者評論文字，計算書本在不同主題的機率。 在實驗中，我們比較傳統方法協同過濾推薦、內容推薦方法單及我們提出的混和推薦方法，經實驗分析，我們發現加入影響力的推薦方法在我們實驗中有較好的推薦結果。

With the increase in use of the Internet, more and more users share, disseminate and search information through the Internet. The Internet provides users an easy way to access a lot of resources (document, photos, media, etc); on the other hand, it causes information overload problems. There is so much information on the Internet that users are hard to find what they really need. In this information overload environment, recommender system becomes important. It uses science method to cope with information overload problem. However, most of researchers utilize either users’ preference or the content of items. The interpersonal influences are lack to be considered; hence, in this paper, we proposed a hybrid recommendation method. It combined traditional collaborative filtering and content-based filtering. Beside, we take interpersonal influences into account. In the collaborative filtering, we add three different influences: review influence, social influence and followed influence. In the content-based filtering method, we use LDA to analysis book review contents and build book topic profiles. In our experiment, we compare our hybrid method with traditional collaborative filtering (UCF) and content-based filtering (CBF). The experimental results show that methods combining all the influences can improve the quality of recommendation.