2.4.1 Clustering aspect
clustering algorithm summed up in about three categories:
A) Simple clustering method according to the principle of minimum distance;.
B) The combined method of two minimum-distance principle;.
C) according to the criterion function. dynamic cluster;
(1) according to the principle of minimum recording simple clustering:
determining a similarity threshold for a specific problem, comparing the distance with a threshold value among the pattern to the respective cluster center, when the value is greater than the threshold as the pattern to another class heart, is smaller than the threshold value to be divided by a minimum distance to a certain class.
Features: Once the category and class of such centers run algorithms to determine the mode will not change.
Summarize the algorithm: advantage is a simple algorithm. However, the clustering process, the center of the class will not change once determined, will not change once it is determined mode. The algorithm relies heavily on the results from the selected threshold T mode and order of participation classification. If you have prior knowledge to guide the selection threshold T, it can often be more reasonable results. Also consider setting different T and select a different order, and finally the good results were compared.
(2) The combined method of two minimum-distance principle (hierarchical clustering method):
First, depending on the respective modes sui generis, then the combined two to a minimum distance-based, this process is repeated, until a date two .
Features: This type of algorithm, like heart constantly revised, but soon after a specified category mode is not changed, is the mode once classified as a class is not being cut open, these algorithms also known as hierarchical clustering method.
(3) dynamic clustering method based on the criterion function:
setting a control parameter of some classifications, define a clustering criterion function can be characterized merits of the clustering process is to make the extreme value of the criterion function optimization process.
Features: The algorithm during operation, like heart constantly revised, category specified in each mode has also been corrected. Such algorithms have --C average method, ISODATA method.
Dynamic clustering basic steps:
1. Establish initial cluster centers, an initial cluster;
category class and distance calculation mode 2, the adjustment mode;
3. calculate the parameters of each cluster, delete, merge or split some polyvinyl class;
4 starts from the initial cluster, using an iterative algorithm to dynamically change the category and the cluster center so that the pattern or criterion function attains its extremum parameters set stop meet the design requirements.
C- here to mean an example method:
1. Conditions and conventions:
pattern feature vectors is provided to be sorted is set: {x1, x2, xn} class number C is determined in advance.
2. Methods The categories and select a given C C initial cluster center distribution according to the principle of the minimum distance for each mode to the class C in a class, then the class is calculated continuously adjust the heart and categories of patterns, eventually each sentence is a mode to its category and the distance from the center square of the smallest.
Example:
known 20 samples, wherein each sample has two, data distribution as shown below, to achieve sample classification (C = 2) using C- mean method;
