傳統網路社群的信任與名聲管理機制相當簡單，通常是由使用者或管理者給予服務提供者評分，會員的可信度與其所得之分數呈正相關。這種簡單機制對於惡意(malicious)、不誠實(dishonest)或挑釁(troll)的會員，並無法提供有效的管理功能，更可能受到惡意成員的扭曲或利用，影響網路社群的正常發展。為解決上述問題，本研究綜合節點分支度、中心性等多種社群網路分析指標，運用線性迴歸與類神經網路，分別以線性與非線性方式加以組合，期能提供正確的社群網路成員影響力與信任度。其次，我們亦考量影響力之極性(polarity)，期能進一步標示出社群中之好群體與壞群體。此外，面對大型社群網路，各種指標的計算往往曠日廢時，本研究亦探討利用不同的中心性指標，預測運算成本較高之指標的可能性。本研究由PTT討論區下載發文資料進行實驗，結果顯示本研究提出之影響力預測方法與實際值具有高度相聯。其次，約有超過70%的成員在社群中的影響力排名預測，誤差不超過25%，上述結果顯示本研究提出方法之有效性。我們也利用四種不同中心性指標來預測中介中心性，結果呈高度相關。此結果對於社群網路分析之成本縮減，提供另一種可行之道。 Trust and reputation management mechanism is quite simple in the traditional Internet community. The feedback score of a member is usually given by other users or site managers. And, the amount of score is positively correlated with the credibility of the members. Such a simple mechanism is unable to provide effective management for malicious, dishonest or troll members but more likely to be distorted or exploited by those abnormal members, which then affecting the normal development of the Internet community. To solve these problems, this research integrates a variety of social network analysis index of degree, centrality, etc. Using linear regression and Artificial Neural Network to combine these indices by linear and nonlinear manner separately. And, use the result to provide correct influence and trust of social networks members. Secondly, we also consider the influence of the polarity, marked a further stage in the community of good groups and bad groups. In addition, the face of large-scale social network, to calculate various index often waste time and drawn out. This research also explores examine the possibility of using different centrality indices to substitute that of high computation cost. Data download from PTT forum are used for experiment, and the results show the effectiveness of the proposed prediction methods. Secondly, the results also show that the prediction error of influence ranking is less than 25% for more than 70% community members. It demonstrate the effectiveness of the proposed method. We also use four different centrality index to predict the betweenness centrality, and the results were highly correlated, which provides another viable way for reducing the cost of social network analysis.
