Davies Bouldin index returns negative values ??

MarcosRL

hello friends community
the metrics applying validation of clusters is obtained and the Davies Bouldin index negative returns me this is a bug?
regards

wessel

As far as I understand, this is not a bug.
Its just easier to use the negative value for optimization.

MarcosRL

excuse my ignorance  , but I do not understand  ???.
In theory lei Davies Boulding index is a number greater than zero.
How's that for optimization you mention?
regards

Stefan_ac

Hey guys,

i am searching for an explaxation to this negative mentioned davies-bouldin values. Please, can anyone explain to me why Rapidminer ist calculation negative values?

 

My Performance Vectors are looking like this:

Performance Vector

Average within centroid distance

cluster_0: -1.831
cluster_1: -1.931
cluster_2: -1.856
cluster_3: -1.897
cluster_4: -1.903
cluster_5: -1.885
cluster_6: -1.891
cluster_7: -1.878
cluster_8: -1.818
cluster_9: -1.869

Davides Bouldin: -1.974

 

Thanks in advance for your help and reply,

Stefan

MartinLiebig

Hi Stefan,

 

i do not know why, but by default the values are multiplied by -1 so that you can run a minimizer on it. That's why the operator has an option called maximize with this description:

maximize

Description: This parameter specifies if the results should be maximized. If set to true, the result is not multiplied by minus one.

 

Simply check it and get what you like more

Best,

Martin

Stefan_ac

Hey Martin, 

thanks for your fast reply.

I read about the multiplication by -1. Thanks for the advanced paramrter advice. Now my values turn in positive ones. BUT, I am still wondering why the values are greater >1. Usualy Davies Boulding values are between 0 and 1 (0="good" clusters and 1="bad" clusters). Now that my values are greater 1, do you have a suggestions for interpretation?

Regards,

Stefan

MartinLiebig

Hi Stefan,

 

why do you think this should be normalized? According to Wikipedia: https://en.wikipedia.org/wiki/Davies%E2%80%93Bouldin_index i don't see any reason to have it in [0,1].

 

Nevertheless you can of course normalize the DB index.

 

~Martin

alle0908

So in either case, does the most optimal cluster according to DB index have the resulting DB that is farthest from zero, or closest? Or asked another way, for the absolute value of DB is a DB index of 10 better or worse than a DB index of 1?

MartinLiebig

Hi,

 

as far as i know smaller absolute values are better. From the doc:

 

davies_bouldin: The algorithms that produce clusters with low intra-cluster distances (high intra-cluster similarity) and high inter-cluster distances (low inter-cluster similarity) will have a low Davies–Bouldin index, the clustering algorithm that produces a collection of clusters with the smallest Davies–Bouldin index is considered the best algorithm based on this criterion.

 

Best,

Martin