Different DBCV value from Matlab's implementation when calculating it on a dataset having all 0s cluster labels

[davidechicco]

Hi Felipe
Thanks again for your availability to fix the issues in your software package during December.
I am writing to you today because I found another case where your dbcv() function generates an outcome which is different from Matlab's implementation one.

I have applied your dbcv() function to this DB1_with_307_clusters.csv
dataset file, where the last column on the right contains the cluster labels, which are all zeros.

I used this piece of code:

import pandas as pd
import dbcv

data_file_name = 'DB1_with_307_clusters.csv'
df = pd.read_csv(data_file_name)
print("data file name read: ", data_file_name)

df = df.drop(df[(df.cluster >= 2)].index)

these_data = df.iloc[:,:-2]
these_clusters = df.iloc[:,-1]
this_dbcv = dbcv.dbcv(these_data, these_clusters, check_duplicates = False)
print("FelSiq/DBCV = ", this_dbcv)

The result is 0.999 and is clearly wrong. A collaborator of mine applied the function of the original DBCV Matlab implementation and obtained 0 as result.

Can you please investigate this problem?

Thanks

-- Davide

[FelSiq]

Hi, Davide.

Could you please elaborate on the expected semantics of the label 0? Is it supposed to represent all noise or a legitimate cluster?

Please keep in mind that the matlab implementation considers "0" the label noise, thus the result being 0.0 makes sense, whereas my implementation considers - by default - "0" a legitimate cluster and "-1" representing the noise labels, thus the result being essentially 1.0 also makes sense.

Am I missing something?

Best regards,
Felipe.

[davidechicco]

Hi Felipe

Thanks for your reply. My file has the following meaning:
-1 noise cluster
0 first cluster
1 second cluster

We are aware of the Matlab implementation's differences, and therefore we added +1 to our data when dealing with Matlab code. We changed the clusters this way:
-1 --> 0 noise cluster
0 --> 1 first cluster
1 --> 2 second cluster

So, our Matlab code does not interpret our 0 labels as noise, but as elements of the first cluster, as it should.
We believe your Python function should produce 0 as result, just like Matlab's implementation
Can you please look into this issue?
Unfortunately I do not have the Matlab code to share with you because I do not have a Matlab license.

Thanks

-- Davide

[FelSiq]

Understood. I'll investigate this asap.

Thanks for the report.

[davidechicco]

Thanks Felipe.
While I'm in it, here's another case where we get different results from Matlab.

In this dataset file (DB1_with_774_clusters.csv), that has the same structure of the DB1_with_307_clusters.csv file mentioned earlier, we get:

• DBCV = 0 with your Python code;
• DBCV = 0.880161 with the Matlab implementation.

Perhaps you want to study it a bit as well.

Muito obrigado! :-)