隨著電子商務網站的普及化建立，消費者或顧客可以透過推薦系統的建議方便購物，而目前以協同過濾的方式來建構推薦系統為市場上的主要技術，為了廣泛收集消費者或顧客的意見與資料，推薦系統必須建構在網際網路環境中，此網路開放的特性同時也使得推薦系統比較容易遭受心懷不軌的人士操弄攻擊，利用灌入一些不公平的假評分資料來扭曲推薦系統所所提供的產品推薦排名，而達成其攻擊的目的。近年來，為了解決推薦系統被攻擊的問題，已經有一些偵測攻擊的方法被提出，但是這些方法在應用上都有一些限制，由於在推薦系統中經常擁有大量遺失值(Missing Values)，以主成分分析(Principal Components Analysis, PCA)為理論基礎的偵測方法必須事先加入大量替代值，此舉會使得整個推薦資料庫不真實；另外，以分類法(Classification)為基礎的偵測技術需要數量相當的真實資料與攻擊資料來訓練分類器，否則會造成分類不平衡(Imbalance)的問題，而實務上收集真實的攻擊資料並不容易；而以統計製程管制(Statistical Process Control, SPC)為基礎的偵測技術，必須在偵測之前取得被攻擊項目之評分機率分配，且仍無法完全移除攻擊者所產生的相關評分資料。為此，本研究在不需要前述方法的限制條件下，我們提出一個βP方法來消除這些被假評分的攻擊問題，而βP方法的理論基礎為統計上的Beta機率分配，βP方法除了容易計算，且透過實驗顯示此方法在MovieLens的推薦系統上對於攻擊資料有著非常穩定有效的過濾效果。Recommender systems are widely deployed to provide user purchasing suggestion on eCommerce websites. The technology that has been adopted by most recommender systems is collaborative filtering. However, with the open nature of collaborative filtering recommender systems, they suffer significant vulnerabilities from being attacked by malicious raters, who inject profiles consisting of biased ratings. In recent years, several attack detection algorithms have been proposed to handle the issue. Unfortunately, their applications are restricted by various constraints. PCA-based methods while having good performance on paper, still suffer from missing values that plague most user-item matrixes. Classification-based methods require balanced numbers of attacks and normal profiles to train the classifiers. The detector based on SPC (Statistical Process Control) assumes that the rating probability distribution for each item is known in advance. In this research, Beta-Protection (βP) is proposed to alleviate the problem without the abovementioned constraints. βP grounds its theoretical foundation on Beta distribution for easy computation and has stable performance when experimenting with data derived from the public websites of MovieLens.
