reset password
Author Message
p0941
Posts: 95
Posted 16:14 Mar 09, 2011 |

1)On clustering,  is it possible for a higher correction rate for a bigger SSE?  It's supposed that a smaller SSE should be more correct.

2) We don't have to do postprocessing for clustering in hw4, do we?  

Thanks

cysun
Posts: 2935
Posted 16:28 Mar 09, 2011 |

1) There's no such thing as a "correction rate" in clustering. The "accuracy" we defined in the homework is, roughly speaking, a measure of how similar the clustering is to the original classification. If they are similar, great; but if they are not, it doesn't necessarily mean the clustering is wrong - it may be that the classification cannot be decided by these attributes.

2) No. Since K is small, we can deal with the initial centroid selection problem by repeating the basic K-Means a number of times.

p0941
Posts: 95
Posted 16:55 Mar 09, 2011 |

Let me modify my question: is it possible for a bigger SSE to have higher accuracy?  Or, does smaller SSE always have higher accuracy when clustering on the same data? 

cysun
Posts: 2935
Posted 16:56 Mar 09, 2011 |
p0941 wrote:

Let me modify my question: is it possible for a bigger SSE to have higher accuracy?  ...

Yes.