When should I use classical k-means clustering and when should I use trimmed k-means clustering?

Question

I suspect that if there are many unimportant outliers, trimmed k-mean clustering should be employed.

Am I on the right track?

Answer 1

Yes, you are correct, trimmed k-means (e.g., Cuesta-Albertos, Gordaliza, & Matrán 1997) is most appropriate when outliers exist and you wish to downweight them during the clustering process. Of course, be cautious and use domain knowledge to determine the appropriate percentage of data to trim as this can impact the results and interpretability of the clusters. Personally, when data is noisy and/or contains outliers I apply several methods such as traditional k-means and k-median clustering.

References

Cuesta-Albertos, J. A., Gordaliza, A., & Matrán, C. (1997). Trimmed $ k $-means: an attempt to robustify quantizers. The Annals of Statistics, 25(2), 553-576.