本研究目的在於解決一般反垃圾郵件軟體(Anti-Spam)未能考慮到個人實際需求，使系統管理員在無法了解個人偏好的情形下，設定垃圾郵件的過濾條件造成正常郵件誤攔的情形，因而設計一種支援回饋機制的個人化垃圾郵件過濾系統，簡稱為PSMF系統(Personalized Spam Mail Filtering System with User Feedback Mechanism)。 PSMF的做法，首先以Lotus Domino建立使用者回饋機制，然後藉由反垃圾郵件軟體所定義的關鍵字，建立關鍵字字庫，最後利用全文檢索的技術，將使用者回饋的垃圾郵件，依照個人的喜好做分類，並將之設定於反垃圾郵件軟體的過濾機制。 本研究已針對實際運用電子郵件處理日常工作的企業員工，實地進行為期兩個月的實驗。以同一組成員分別進行實驗與對照的觀察，每日紀錄攔截與誤攔郵件的比例，藉由本論文所設計的回饋機制與過濾規則的設定，實驗結果證實，實驗組的攔截率高於對照組，而誤攔率的結果也比2003年五月號PC World 雜誌評比結果的平均值17%低，顯示經由使用者實際回饋的機制作為過濾依據，確實較能提高攔截率與降低誤攔的比例。 This paper is focus on solving the problem that Anti-spam software faces. It did not consider the user’s preference, which caused false negatives and false positives because the system administrator cannot understand the preference of users. Therefore, a Personalized Spam Mail Filtering System with User Feedback Mechanism (PSMF) is designed to setup private filtering rules to match users' preferences. PSMF embedded a feedback program into users’ mailbox to collect their abominated emails. We categorized spam mails with the full text search technology of Lotus Domino according to the classification defined by anti-spam software. Finally we setup the filtering rules on the anti-spam software for our users. The function of PSMF first adopted the full text search technology of Lotus Domino according to the classification defined by anti-spam software. Then, a feedback program was embedded into users’ mailbox to collect their abominated emails. Finally, the filtering rules were setup on the anti-spam software for our users. An experiment on PSMF system was conducted for two months on employees of a corporation who frequently uses the email. Two sets of tests was done with the same group of employees based on two conditions; the first test was to use the PSMF system for a period of time and the second test was not to use the PSMF system for a period of time. The result of the two tests were compared and it showed that PSMF system applying the private filtering rules has a higher succession rate and achieve less false positives than the average rate of 17% of an experiment conducted and published in May 2003 issue of PC World.
