臺大管理論叢 NTU Management Review VOL.28 NO.3

Use of Implicit User Feedback to Support Semantics-Based Personalized Document Recommendation 100 Table 8 Analysis Results of Mixed Model ANCOVA (Three Treatments) Source Sum of Square df Mean Square F Sig. Recommendation Technique 15.9161 2 5.3054 8.8936 0.0000 Stage 0.0014 1 0.0014 0.0091 0.9243 Top 3 rating 14.9416 1 14.9416 25.0472 0.0000 Treatment x Stage 0.2096 2 0.0699 0.4395 0.7253 Within-Group 51.945 132 Subjects 42.125 66 0.638 Error 9.820 66 0.149 Total 83.0137 138 Note: The random-based group was excluded. 5. Conclusion This study proposed an Implicit-feedback-based Concept-Expansion (ICE) document recommendation technique to address the limitations inherent in the acquisition of users’ preference feedback. The ICE observes and analyzes users’ browsing behaviors implicitly to assess their interests, and identifies extra concepts related to the focal user’s document interests in order to provide appropriate, personalized document recommendations. The results of our empirical evaluation suggest that our proposed ICE outperforms both keyword-based and random-based recommendation approaches. Although the performance difference between ICE and ECE is not significant, the ICE intrinsically requires less user effort than does the ECE. Our evaluation results also match the findings of a prior study in that implicit-feedback-based and explicit-feedback-based collaborative filtering recommendation techniques were found to perform about the same (Jawaheer et al., 2010). The implications of this study are as follows. First, our proposed ICE technique determines the user’s interests based on browsing time rather than preference rating. This suggests that users’ interests can most likely be gathered or learned implicitly from their browsing behaviors. Content providers can make better recommendations without users’ direct feedback, or at least with less explicit feedback, which usually requires constant user input and is not easy to collect. In other words, our approach requires less extra effort from users, thus enhancing their satisfaction and increasing the value of the firm (Chen, Lin, and Tai, 2011). In particular, some explicit-based recommender systems require users to provide their preferences by keying in data. However, some disabled people cannot easily type in such data, which may stop them from enjoying the benefits of the