PNN(Pairwise-nearest-neighbor)是一種有效的資料分群方法，且通常可以產生良好 的分群結果，但是計算複雜度很高。目前已有許多快速的PNN 方法可用來改善PNN 的計算複雜度，在這些快速的PNN 方法中，Fränti 等人所提的方法(簡稱為FPNN_LPM) 與我們所提的方法(MFPNN_FS)是目前可用方法中最快的兩個方法。不過，這兩個方法 很容易受到群組分離度的影響，當資料集的群組分離度低時，使用FPNN_LPM 與 MFPNN_FS 方法改善PNN 的效果不好，也就是說，這兩個方法只適合高群組分離度 的資料集使用。這個問題的主要原因出在搜尋起點的選擇方法以及過濾條件不適用低 群組分離度的資料集。為了解決這個問題，本計劃將探索群組分離度如何影響搜尋起 點的選擇方法以及過濾條件的表現，然後開發出適合各種群組分離度的可調適性的搜 尋起點選取方法以及可調適性的過濾條件來改善PNN 的執行效能。 Pairwise-nearest-neighbor (PNN) is an effective method of data clustering, which can always generate good clustering results, but with high computational complexity. Many fast exact PNN methods are proposed to reduce the computational complexity of PNN. Among available fast PNN methods, two methods respectively proposed by Fränti et al. (referred to as FPNN_LPM) and me (referred to as MFPNN_FS) are two best fast exact PNN method. However, both methods are very sensitive to the degree of cluster separation. For a data set with low degree of cluster separation, the performances of applying FPNN_LPM and MFPNN_FS to improve PNN are poor. That is, these methods are suitable only for a data set with high degree of cluster separation. The major reason of this problem is that the initial searching point selection method and rejection criteria are not suitable for low degree of cluster separation. To solve this problem, we will discover how the degrees of cluster separation affect the performances of initial searching point selection method and rejection criteria and then develop an adaptive initial searching point selection algorithm and an algorithm with adaptive rejection criteria to improve the performance of PNN, which are suitable for any degree of cluster separation.
