結合多面向信任及協同過濾之部落格文章推薦系統

Abstract

網際網路的蓬勃發展，為人類帶來前所未有的便利性，卻也因此而產生了資訊超載的問題，而推薦系統的出現，有效降低了人們判斷具價值資訊所須耗費的成本。部落格是近幾年網際網路上的新殺手級應用，其賦予網路使用者在網路上表達自我並分享知識的管道，而每年的新部落格數目更以指數速度成長。由於部落格作者本身撰寫文章主題的歧異性和變化性，使得一般網路使用者往往難以有效的從數以百萬計的部落格文章中，獲取對其來說真正具閱讀價值的文章。目前部落格推薦的網站大都只提供部落格的搜尋服務，而無法提供個人化的部落格文章推薦來協助網路使用者減輕資訊超載所造成之負擔。 有鑒於此，本研究修改過去文獻之作法，提出一整合使用者之間對不同文章種類之多面向信任資訊，與使用者相似度的協同過濾演算法，並實際實作了一個線上部落格文章推薦系統 – iTrustU，以驗證本研究提出之方法是否能夠有效改善推薦系統的準確度及滿意度。由179位網路使用者(部落客/一般閱讀者)所進行為期45天的線上實驗中，我們發現不論是在推薦準確度或是使用者滿意度上，本研究之系統皆獲得非常高的評價。相較於只單獨考量相似度或信任資訊的傳統協同過濾演算法，本研究提出之整合性演算法顯著地具較高的準確度，尤其改善了對於冷初始(cold start)使用者的推薦品質。而透過統計分析，我們進一步證實在部落格社群中，使用者彼此之信任與相似度之間的確具有顯著的正相關，此研究結果與過去研究相互呼應。本研究之研究結果有效地證明，透過使用者信任網路中信任關係的運用及推演，可協助推薦系統提供更好的推薦服務，而我們所提出的整合性協同過濾演算法，並不僅限於部落格社群，其也可適當地應用於任何存在使用者之社會信任網路資訊的領域之中，如線上社群網站、購物拍賣網站等等。

The evolution of the Internet has given people access to information in a way never previously imagined; yet, ironically, it has given rise to the problem of information overload. Fortunately, the advent of recommender systems has relieved people of much of the effort required to find desired information. Blogs represent a new killer application on the Internet that gives users a channel to express themselves and share their knowledge and feelings with other people worldwide. The number of new blogs is growing exponentially. However, due to the diverse subjects covered by bloggers, it is difficult for readers to find blogs containing articles that fit their interests or information needs from the hundreds of thousands, possibly millions, of blogs on the Internet. Currently, most blog recommendation websites only provide search functions based on different types of blogs. In other words, they do not provide any customized or personalized blog article recommendations. Given the need to ease information overload in the blog domain, we have modified some existing approaches, and herein propose a novel trust-enhanced collaborative filtering approach that integrates multi-faceted trust based on article types and user similarity. We also designed an online blog article recommender system, called iTrustU to evaluate whether our proposed approach can improve the accuracy and quality of recommendations. During a 45-day online experiment with 179 participants from the Internet, we found that our system achieved good outcomes in both recommendation accuracy and user satisfaction. In contrast to traditional collaborative filtering approaches, which only consider user similarity or trust information, our integrated approach yields a significantly higher accuracy, especially for cold start users. Through statistical analysis, we prove that in the blogosphere community, trust and similarity among bloggers/readers exhibit a significantly positive correlation. This result is the same as that of past research. Our research results show that, through the exploitation and inference of trust relationships in a trust network, we can provide more effective recommender systems in terms of user satisfaction. The proposed approach not only applies to the blogosphere, but also to any online social community or commercial shopping/auction websites when trust relationships already exist between users on the fly.