模式识别讲稿2010-lecture7.ppt

Data Description and Clustering methods. The use is usually justified on the ground that they are easy to apply and often yield interesting results that may guide the application of more rigorous procedure Similarity Measures: Criterion Function For Clustering No. of samples in different clusters should be About the same. ( check the graph in the following slide) Scatter Matrices (similar as defined in the section of Fisher’s linear discriminants) Hierarchical Clustering It provides good starting points for iterative optimiz- ation given n samples. Start with n cluster ━ 1 per sample ------ level 1 next partition into n-1 clusters and then n-2 clusters and so on. at the kth level, there are Go to step 2. classes, the agglomerative procedure becomes one for generating a minimal spanning tree. Fig. 6.16 , Fig. 6.17. one vector will decrease if Transfer split into 2 classes by forming two new class center: where has component each of which is the square-root of If splitting does actually take place in step 10, go to step 2 , otherwise continue. Step 11. (merge classes) Compute the pairwise between-centre distances: used in lumping in this iteration, merge these once 1 iteration Isodata is one of a class procedures known as decision-directed procedure (2)The act of classification will exclude samples from the tails of the true distribution, and will include sample from the tails of the other distribution. * One can virtually approximate any density function by a mixture of c normal distributions, if c is large enough. If we don’t even know the labels of the samples, it’s doubtful about the soundness of own knowledge on and the form of The alternative is to use non-parameter clustering Clustering means finding natural groupings of samples. similarity measure between samples evaluat