依據美國部落格搜尋網站公司Technorati[44]於2007年的報告指出,部落格網站的數量已快速攀升至7,200萬且越來越受到許多部落客的青睞,也因此讓許多部落格網站使用者逐漸地轉型成部落客(Blogger)。然而,隨著部落格資訊的增長,也產生了現今大多數部落格網站無法避免的文章資訊過載(Information Overloading)現象。因此,如何有效協助資訊之查找,如何有效的推薦文章,已成為不可或缺的個人化服務,這也是目前多數的部落格網站所缺乏的。 本研究為能有效推薦符合瀏覽者需求的文章且能克服上述問題,因此,提出一個部落格網站系統平台,命名為BARS(Blog Article Recommendation System),BARS可以用來分析瀏覽者的偏好興趣,並將瀏覽者的偏好興趣紀錄依本體論(Ontology)階層概念建構出樹狀型態,用以表達瀏覽者的歷史閱覽記錄及挖掘潛在興趣偏好,此外,本研究亦利用類神經網路之模型—自適應共振理論網路(Adaptive Resonance Theory, ART)來將具有相同偏好興趣的瀏覽者進行分群,分群後將具有相似鄰居性質的瀏覽者透過推薦系統(Recommendation System, RS)中的協同過濾法(Collaborative Filtering, CF)來找出目標瀏覽者與其他鄰近瀏覽者相似的潛在偏好。針對那些沒有相似鄰居群的目標瀏覽者,則透過推薦系統的另一項方法--內容導向法(Content-Based method, CB),依瀏覽者過去的瀏覽歷史記錄及本體論概念的推論方式來推論出使用者未曾瀏覽過卻可能產生的興趣偏好,藉此亦可降低冷啟始(Cold-start)所產生因起始資料不多而無法推薦的問題。基於上述研究理論,本研究欲達成的目標及貢獻為:(1) 解決因資訊過載、文章資訊搜集繁雜議題,(2) 降低瀏覽者冷啟始問題,(3) 提昇個人化之文章推薦效能,(4) 提出部落格推薦應用之先導研究。本研究依照上述方法已完成系統建置與實驗,本研究運用三種推薦方式並與傳統的非個人化推薦方法(NPRL)做比較,這三種方法為BRL(Browser-based Recommendation List)、CRL(Cluster-based Recommendation List)及SCRL(Spread Category Recommendation List)。實驗結果顯示使用本研究BARS系統所提之BRL方法,其文章推薦命中率可達到84%;在SCRL方法上也可獲得83%的推薦命中率,而第三個方法CRL雖是三個方法中較低的,但仍具有高達80%的文章推薦命中率,上述三種方法都較傳統的NPRL方法之效能更佳。
According to Technorati’s [44] 2007 report, the blog sites is up rising to 72 millions and the popularity of bloggers has drawn many attention. This phenomenon has turned many web users into bloggers. The vast amount of blog information also brings the phenomenon of information overloading which is not handled by the blog function yet. In addition, the personalized recommendation service, which should be provided, is also not incorporated in the blog function now. To better service the bloggers and to overcome the above problems, this research proposes a Blog Article Recommendation System (BARS) which provides personalized article recommendation based on blogger’s preference. This research adopts the ontology technique in BARS to construct a personal preference tree for recording blogger’s interests and for further inference. The ART (Adaptive Resonance Theory) network is also utilized to cluster the group of similar interests. In order to find the similar preference between target blogger and the corresponding neighbors, this research applies the Collaborative Filtering (CF) technique to generate the recommendation. The “cold-start” problem, i.e. lacking of blogger’s usage data at the very beginning, is handled by integrating ontology and Content-Based (CB) filtering method to infer the potential preference in BARS. The purpose of this research is to achieve the followings. (1) To solve the problem of information overloading. (2) To handle the cold-start problem when making the recommendation. (3) To implement BARS for fulfilling the personalized blog-article recommendation. (4) To act as the pioneer for providing recommendation service in blog’s research. To demonstrate the proposed method, this research has implemented the BARS as an academic blog for real-world experiment. Experiments are done by collecting real bloggers’ usage and their feedbacks so that the recommendation correctness is verified. Three types of recommendation approach are proposed and compared with the traditional Non-personalized recommendation list (NPRL), these three types of approach are BRL (Browser Based Recommendation List), SCRL (Spread Category Recommendation List), and CRL (Cluster Based Recommendation List) method. The experiment shows that BRL produces better recommendation result than SCRL and CRL with 84% satisfaction by all the bloggers.