I try to use k-means clusters (using SQLserver + R), and it seems that my model is not stable : each time I run the k-means algorithm, it finds different clusters. But if I set nstart (in R k-means function) high enough (10 or more) it becomes stable.
The default value for this parameter is 1 but it seems that setting it to a higher value (25) is recommended (I think I saw somewhere in the documentation).
So I'm a bit confused... Any advice ?
nstart option attempts multiple initial configurations and reports on the best one. For example, adding nstart=25 will generate 25 initial random centroids and choose the best one for the algorithm. Hope this helps!
Stability of the clusters is highly dependent on your dataset, for clear cut cases running it multiple times is a waste of resources. I think that is the rationale behind the default value of 1. But I agree that for most smaller cases setting it much higher makes a lot of sense.
Simply put, nstart will create multiple configurations and will show the best one.
For example, nstart = 25 , will create 25 initial configurations.
Use nstart argument to try multiple random initial values.
the explanation doesn't appear naywhere in doc. but I think that setting nstart higher let kmeans run nstart(say 50) random initialization of centroids and choose the one that minimize better the cost; therefore you end up with a much more stable clusters becouse of kmeans always choose the better separation for your data
