algorithm - Can k-means clustering do classification? -
i want know whether k-means clustering algorithm can classification?
if have done simple k-means clustering .
assume have many data , use k-means clusterings, 2 clusters a, b. , centroid calculating method euclidean distance.
cluster @ left side.
cluster b @ right side.
so, if have 1 new data. should do?
run k-means clustering algorithm again, , can cluster new data belong to?
record last centroid , use euclidean distance calculating decide new data belong to?
other method?
the simplest method of course 2., assign each object closest centroid (technically, use sum-of-squares, not euclidean distance; more correct k-means, , saves sqrt computation).
method 1. fragile, k-means may give different solution; in particular if didn't fit data in first place (e.g. high dimensional, clusters of different size, many clusters, ...)
however, following method may more reasonable:
3. train actual classifier.
yes, can use k-means produce initial partitioning, assume k-means partitions reasonable classes (you really should validate @ point though), , continue if data have been user-labeled.
i.e. run k-means, train svm on resulting clusters. use svm classification.
k-nn classification, or assigning each object nearest cluster center (option 1) can seen simple classifiers. latter 1nn classifier, "trained" on cluster centroids only.
Comments
Post a Comment