本篇論文提出了以挑選可靠的候選集為基礎來解決探索朋友關係及個人興趣的問題。為了解決上述這些因不完整的個人資料所導致的問題。區域中朋友數最多的節點被拿來建立預測朋友關係的分類器,以還原真正的個人資訊,並且利用其在不同社群中分佈的密度來預測每個人隱藏朋友的比例。在每一回合中,分類結果信心指數最高的候選集才會被執行,進而避免產生不正確的朋友關係及個人興趣。為了評估我們方法的效果,我們與信心指數決定次序的方法及固定推薦次序的方法做比較。結果顯示,在不同的資料集中皆改善了效能,特別是在連結稀疏的網路。
The thesis presents a reliable candidate selection approach for link prediction and interest inference. To deal with the above problems caused by the in- complete personal information, the nodes with local maximal degree are used to build the classifi ers for recovering the hidden information. Moreover, the hidden friendship of person is estimated based on the local maximum density of the community, and filtering the incorrect inference and prediction as well. In each round, the candidate with maximum con fidence of the classi fication will be recovered iteratively to reduce the generation of mistaken interest and friendship. To evaluate the proposed method, we compare the performance with con fidence-based and iterative method. We improve the performance in various data sets, especially the sparse network.