為協助網路使用者從大量文件中取得所需的資訊，文件推薦系統已成為必要的支援工 具。本研究設計與評估以隱性回饋為基礎的概念擴展文件推薦技術(ICE)，其採用隱 性觀察與使用者瀏覽行為分析，來克服擷取使用者回饋的困難，並加以利用領域本體 論與擴散促動模式，擴展已知的偏好文件中的概念，以推薦合適的文件給使用者。此 研究利用實驗室實驗法來評估ICE 系統與另外三個指標系統的差異，包括：顯性回饋 為基礎的概念擴展文件推薦技術(ECE)、關鍵字基礎技術、及隨機推薦系統。實驗結 果顯示ICE 績效顯著優於隨機推薦及關鍵字推薦，與ECE 雖無顯著差異，但是預期 ICE 能降低使用者的心力耗費。整體而言，此研究結果提供改善文件推薦系統品質之 實務意涵。

The development of the Internet and digitized documents has made it possible for data and information to be easily transferred, exchanged and shared online. However, for Internet users, this easy access to information also carries the risk of information overload. Document recommender systems are becoming an indispensable tool, helping Internet users effectively retrieve the information they need from the millions of documents available online. In this study, we design and evaluate an Implicit-feedback-based Concept-Expansion (ICE) document recommendation technique to address the difficulties inherent in acquiring relevant feedback. The ICE technique determines a focal user’s preferred documents by implicitly observing and analyzing his or her browsing behavior in order to make appropriate document recommendations. Using a domain concept heterarchy (e.g., domain ontology) and employing the Spreading Activation Model (SAM), the ICE technique expands the concepts existing in the preferred documents. Documents with a greater number of related and/or expanded concepts are then considered to be potentially appealing to the focal user and are recommended as such. A laboratory experiment was conducted to compare the system performance of the ICE technique with that of three benchmark document recommendation techniques: Explicit-feedback-based Concept-Expansion (ECE), keyword-based, and random. The results of the experiment show that the ICE approach proposed by this study is more effective than random or keyword-based document recommender systems. Although there is no significant performance difference between ICE and ECE, the ICE technique is expected to cost less in terms of user effort. Overall, the findings of this study provide some interesting implications for improving the quality of document recommender systems.