推薦系統目前廣泛的被使用在電子商務環境中，協同過濾演算法(Collaborative filtering algorithm)最常被使用，但協同過濾法有一些缺點，例如是新的使用者、新的推薦項目通常都沒辦法獲得準確推薦、資料庫稀疏的問題、以及惡意評比而影響推薦準確度，本研究改進協同過濾推薦的惡意評比問題，在方法上，提出在協同過濾演算法之前的過濾機制是以beta分配為基礎，此方法通常運用於線上拍賣系統中，本文藉此過濾機制以預測是否此機制可以有效的刪除惡意攻擊推薦系統的推薦準確度。實驗所使用的資料是以MovieLens的資料庫為主，在本實驗中把使用者按照電影種類來區分，個別觀察其推薦的項目、次序。 實驗共分成兩個，第一個是混入人工製造的資料，此可視為惡意攻擊之評比；第二個是用原本的MovieLens的資料庫的資料，因此在實驗評估共分兩個，第一個是比較加入人工製造之資料是否會影響推薦的項目。第二是比較是否在兩者實驗刪除部分使用者後其推薦項目是否會被影響。本實驗結果是在實驗二，用MovieLens原始資料對刪除部分使用者前後的推薦項目並無太大差異，且在混入人工製造資料的實驗中過濾偽評比後可以完全刪除人工製造之評比外，其推薦項目和在使用原本的MovieLens的資料庫的評比尚未做刪除機制的結果亦無太大差異。 Recommender systems are widely used in the e-commerce environment. Until recently, most used algorithms or collaborative filtering methods; but in these approaches, shortcomings: new use and item problem, as well as sparsity and attacks on recommender systems that will affect its precision. We propose a solution to solve problems caused by attacks on recommender systems using a beta reputation system in order to filter out unfair ratings in the recommender systems. This method is popular in the on-line marketplace but we have improved the ability to delete unfair rating existing in the recommender system though this filtering method. The dataset we used is MovieLens. We group users’ ratings by movie categories and observed the difference of recommendation items. There are two kinds of data content in our research; the first one is combines the MovieLens dataset with man-made ratings which can be seen as ratings by attacker. The second one is original MovieLens dataset. There are 2 evaluations: average difference in percentage of two experiments and average deleting users of two experiments. The advantage in our proposed method is that it filters out fake users in the recommender system. The results show that the accuracy and the performance are considerably improved.
