.

could you give a little more details (or examples) of what you call "complementary" objects/patterns ?

.

For instance, there are two wind turbines: Wind turbine A located has abundant wind during the first half of the day i.e. from 1 am to 12 noon and produce nothing in second half of the day i.e. from 13 to 24 hours. Whereas wind turbine B has opposite pattern means it has nothing on the first half of the day but has abundant of wind on the second half of the day. i am looking for some clustering algorithm which take consideration of those pattern to make clusters.

.

interesting ; i never considered such problem ...

i guess i would define some similarity between objects in terms of **negative** correlation (the strongest the negative correlation between objects, the most similar they are) and run any kind of clustering on top of that similarity ... as a first guess

.

edit :

this might work only with single link hierarchical clustering ; clusterings based on centroids will spoil everything (once i have a cluster with two negatively correlated items , if i average them, i recover some shapeless pattern !)

single link hierarchical clustering might do the job in the sense that an item joins a cluster if it is most strongly negatively correlated to one of the members of this cluster (one to one comparison makes sense and keeps the constraint of addind items negatively correlated to some items in the cluster

.

edit bis :

googling around, i found this :

http://www.marcottelab.org/paper-pdfs/TechnicalReport_DiametricClustering_tr02-49.pdf

(i have'nt read it yet but it argues that a clever (spectral) choice of the cluster representative makes agglomerative clustering possible, at variance with what i wrote just above ! Thanks for giving me the opportunity to consider this problem !)

.

Thanks for your detailed answer. Coming back to your idea of taking correlation and making clusters with those having highest negative correlation. Don't you think that this approach is not computationally tractable?

.

why would it be more intractable than standard single link clustering (which indeed poses severe challenges for very large databases) ; do you study gazillions of turbines ?

.