0
$\begingroup$

I would like to classify individuals of a database by a single quantitative variable. Is hierachical clustering suitable to do this? If it is possible, how the algorithm work? If use hierachical clustering is not right, which other procedures or techniques could I use? My purpose is not classify the variable, but individuals. So decision trees are not suitable.

$\endgroup$
0
-1
$\begingroup$

You could order the nob values from low to high and then use a search procedure to identify when and if local mean(s) changed significantly via Intervention Detection (trial and error). ID is esseentially a single dimension(characteristic) cluster analysis. Alternatively you could pre-specify the number of groups (classes) that you wished to have(n) and then find the n-1 breakpoints which optimally classifies the nob values. I have not ever done this but it might be worth a try.

$\endgroup$
6
  • $\begingroup$ Sorry, I don't understand your approach. What is nob value? Could you explain your approach in more details? Can I run a hierachical clustering? Thank you. $\endgroup$ – Jos Oct 15 '13 at 15:23
  • $\begingroup$ As usual, @IrishStat, your inclination is to treat everything as a time series! If the data are not a time series, I think the validity of your method depends on whether the same classes (breakpoints) would be identified from a series and itself reversed, i.e. the method depends on past and future being interchangeable. $\endgroup$ – Nick Cox Oct 15 '13 at 15:43
  • $\begingroup$ I am not working with time series... What methods can I use? Can I use multiple classification methods (like hierachical ascending) by only one variable as a classifier? $\endgroup$ – Jos Oct 15 '13 at 15:49
  • $\begingroup$ @Nick The trick when using Interevention Detection for non-time series data (whis is what we have) (successful I might add) is to disable ARIMA identification, seasonal pulse identification and trend detection. These constraints eliminate any and all unwarranted/unwanted time series structure. All that is left to identify is level/step shifts (group classification) and pulse detection (one time anomalies). Send me your email address and I will forward you the results of any set of values that you wish to send to me. $\endgroup$ – IrishStat Oct 15 '13 at 16:20
  • $\begingroup$ @IrishStat Thanks, but I already have code that does something similar to my satisfaction (I wrote the answer to the thread cited as duplicate to this). I was just curious about any hidden assumptions behind your suggestion. $\endgroup$ – Nick Cox Oct 15 '13 at 16:36

Not the answer you're looking for? Browse other questions tagged or ask your own question.