I need to know each point to which cluster in kmeans fun.

hi, i used this function kmeans, and take simple example

 x=[100 2 4 10 200; 50 100 20 1 5];

[p o]=kmeans(x,2)

what i got:

p =

     1

     2

o =

   100     2     4    10   200

    50   100    20     1     5

what i need is know each point for which cluster belong, what the code do is just get the two clusters i gave it and the points that i gave it.

how I know the clusters that each point belong to?

thanks

[huda nawaf]

to be more clear, if I have array : x=[100 2 4 10 200; 50 100 20 1 5; 3 4 5 6 6;....]; this array show the rate of similarity. where the first row show the similarity rate between user1 and the all other four users, the second row show the rate of similarity between second user and the others and so on. how I know each user for which cluster belong?

thanks

[Yash]

try to plot the result also

Answer 1

[Peter Perkins]

Huda, you gave kmeans two points and asked it to cluster them into two clusters. It has assigned the first point to cluster 1, whose centroid is at the first point, and similarly for the second point. Presumably that is not very informative.

I don't know what your data mean, or whether kmeans makes sense, but your description sounds like something more suited to distance-based methods such as hierarchical clustering or multidimensional scaling, both of which are available in the Statistics Toolbox. You would, of course, have to convert your similarities to dissimilarities.

[huda nawaf]

thanks,

i used statistic toolbox

X = [0 4 6;2 0 4;3 5 0);

each row represent rates of similarity of one user with the other, what i need is clustering of users

x=magic(3);
c = clusterdata(X,'linkage','ward','savememory','on',...
    'maxclust',4);
scatter3(X(:,1),X(:,2),X(:,3),10,c)

still ,there is something is not clear, how I know the cluster of each user?

[Walter Roberson]

The output of clusterdata is the cluster number for each sample.

Answer 2

[Walter Roberson]

The first output from kmeans() is the cluster number for each sample.

http://www.mathworks.com/help/toolbox/stats/kmeans.html

IDX = kmeans(X,k) partitions the points in the n-by-p data matrix X into k clusters. This iterative partitioning minimizes the sum, over all clusters, of the within-cluster sums of point-to-cluster-centroid distances. Rows of X correspond to points, columns correspond to variables. kmeans returns an n-by-1 vector IDX containing the cluster indices of each point. By default, kmeans uses squared Euclidean distances. When X is a vector, kmeans treats it as an n-by-1 data matrix, regardless of its orientation.

[huda nawaf]

thanks, you are meaning that the output will be the no. of cluster for each row, because the points corresponde to rows. right? i.e when x=[1 2 3; 4 5 6;]; [p o]=kmeans(x,2) p=1 2 is that meaning kmeans classify each row to spesific cluster? why when use x=[1 2 3]; p= 1 1 2 x now is one row.

[huda nawaf]

for more clarifying, I have 1000 users, I found the similarity by depending on thier features among them using SEQUENCE ALIGNMEENT. Then, I got 1000*1000 matrix of similarity. What is the best way to cluster users to find communities? kmeans or clusterdata

[Peter Perkins]

Please read my previous comment.

[huda nawaf]

I read it, this is why I used clusterdata, because I need hieraricial clustreing. but I don't understand how clusterdata works.

I tried this example if i hve three users,the similarity among them as user1 0 10 20 user2 10 0 5 user3 20 5 0

where zero is the rate of similarity between any user with himself. when I used: T = clusterdata(x,'maxclust',3)

T =

     1
     2
     3

I was expecting that user1will be in cluster 2, belong cluster 3,and user3 belong to cluster2 depending on least distance. in the output above ,I don't why cluster the first user for cluster1 and so on. thanks

[Walter Roberson]

I do not mean that the output will be the number of clusters for each row. I mean that the output is will be the cluster number for each row. "number of clusters" would be "how many clusters occur in that row". "cluster number" is the index of which cluster the row was assigned to.

The cluster numbers have no separate meaning. It does not matter whether the order is "cluster 2 then cluster 3" or "cluster 3 and then cluster 2": the important part is consistency and that the cluster centroid is the right one (you are not outputing the cluster centroids.)

[huda nawaf]

sorry , may be im not clear. I;m also do not mean the no. of clusters. what I mean , I thought that user1 belong to second cluster, user2 belong to third cluster, and user3 belong to second cluster depending on least distance. where each row represet the rate of similarity with other users.

in this case , is clusterdata suitable way for clustering? thanks

[Peter Perkins]

Walter, huda is using hirarchical clustering now, not k-means, so I think he was responding to my earlier post.

huda, again, you have given clusterdata three points,andf asked for three clusters, and it has returned what you asked for. The cluster number is completely arbitrary. I suspect that if you used linkage, then dendrogram, you'd see what you were expecting

d = [0 10 20; 10 0 5; 20 5 0] % d is a distance matrix
z = linkage(squareform(d)) % convert d to vector form first
dendrogram(z)

The dendrogram plot has labels on its horizontal axis that refer to the original points, not the cluster numbers).