群體決策問題在最近幾年中，因為被應用到多種的領域上而備受注目。群體決策問題主要是從收集到的使用者喜好資料當中，找出共識的序列代表這些使用者的喜好以利決策者做決策。在過去的研究中，有完整排名序列及最大共識序列兩種輸出格式，兩者各有其優點而被廣泛採用，然而完整排名序列雖然藉由匯總使用者喜好資料的方式，建立了一組所有項目的序列結果，但是沒有考慮到有多少喜好衝突存在。另一方面，最大共識序列克服了完整排名序列的缺點，提高了結果的共識程度，也指出喜好衝突的部份待作進一步分析，但是零散的序列結果並不容易被理解及使用。為改善以上兩者的缺點，我們嘗試利用對使用者喜好排名序列分群的方式，找出所有項目的共識群。共識群的優點是群中可以容忍喜好衝突存在，而且結果也容易被理解及使用。我們的研究定義了兩個群定義，透過分割式分群演算法的k-mean及k-medoids的概念進行兩個階段的分群。最後，我們透過不同參數設定進行實驗分析並作討論及總結。 The group ranking problem is used to obtain the coherent result from users’ preference data and has received increased attention due to its widespread applications in recent decades. Among previous researches, two output formats have their own advantage and have been used in many applications. Nevertheless, total ranking list consolidates a consensus ordering list by aggregating users’ preference data regardless of how many conflicts exist. On the other hand, maximum consensus sequence has overcome the degree of majority and identified the conflicts that need further negotiation. However, the results of maximum consensus sequence are fragmented, and not easy to understand and use. To conquer these disadvantages, we try a novel way called consensus clusters to cluster user ranking lists. According to two cluster definitions we defined, our methodology is developed based on k-means and the concept of its variant k-medoids to minimize the cost functions. These consensus clusters tolerates the conflicts of item preference and is easy to understand. At the end, we discuss our experimental results and contributions.
