結合多準則評分信任與協同式過濾之文章推薦

Abstract

協同過濾推薦系統已經廣泛地被應用在各種商業領域解決資訊過載等問題。近年來，已有許多的推薦系統開始採用多準則評分(multi-criteria ratings)，希望藉由讓使用者在物件的不同面向上給予個別的評分，能夠更深入地瞭解使用者的評分行為與喜好，進而找到喜好更為相近的鄰居，並提升推薦的品質。目前，多數有關多準則評分推薦系統的研究集中在討論預測整體評分(overall rating)，或是以追求整體評分效用最大化為其推薦的策略，然而有些多準則評分的推薦系統並沒有提供整體評分，此時，缺乏使用者的最後決斷資訊(整體評分)讓推薦的程序變得相當複雜。同時，在某些情況下，也不適合採用追求整體評分效用最大化的推薦策略，例如：當評分準則之間存在衝突而無法相互取代。本研究整合多準則評分與傳統信任推薦，提出一個多準則評分信任的推薦方法，以及一個混和(hybrid)方法，此混和方法整合多準則評分信任與協同式過濾等推薦技術。同時，針對上述問題，本研究提出一個2階段的推薦程序：(1)、設定一組推薦門檻，將之當成推薦過濾器，(2)、分別發展三種不同的推薦策略。實驗結果顯示，本研究的方法能在沒有整體評分的多準則評分推薦系統中，提升推薦的品質。

Collaborative filtering recommender systems have proven to solving information overload problems, and are widely implemented in various industry domains, such as books, music, video rentals, hotels, and so on. Nowadays, many recommender systems are extended from single-criterion rating to multi-criteria rating systems in order to improve the quality of the recommendations by selecting more similar neighbors based on multi-criteria ratings. The researches related to multi-criteria rating recommender systems mainly focus on predicting overall score and pursuing maximal utility for producing recommendation lists. However, some multi-criteria rating recommender systems do not have overall rating. Accordingly, producing a proper recommendation list in such systems become challenging due to a lack of final judging (overall rating) of their users. In some circumstances, such as conflicts among criteria, pursuing maximal utility could not be treated as best strategies because some criteria could not be replaced with others. In this paper, we incorporate multi-criteria ratings into the conventional trust-based techniques and propose a hybrid model combining multi-criteria rating-based trust and collaborative filtering techniques. To eliminate the problems mentioned above, we propose a 2-step recommendation process: (1) setting a set of recommendation conditions as recommendation filters, and (2) applying three recommendation policies to recommend items. The experiments show that our proposed approaches can increase the recommendation quality of multi-criteria rating recommender systems without overall rating.