dm.cs.tu-dortmund.de/mlbits/cluster-kmeans-limitations/
Limitations of k-Means Clustering – Lecture Notes
\smash {\mu _S-x}\mathstrut \right\rVert ^2 \end{align*}\]
It can be better (change \(\Delta \lt{}0\) ) to not assign points to the nearest cluster!
Using \(\textup{SSQ}(A\cup B)=\tfrac {|A|\cdot |B| [...] Initially, let point \(p\) be assigned to cluster \(S\) , which yields a sum of squares of 9+4+4+9+0=26.
Assigning the point \(p\) to \(T\) will yield the following situation, with sum of squares 4+5+5+4+4=22 [...] \(\textup{SSQ}\) for different \(k\) or different data (including normalization)
\(\textup{SSQ}_{k=N}{=}0\) — perfect solution? No: useless
\(\textup{SSQ}_k\) may exhibit an elbow or knee : initially it improves …
