[dbscan] [enhancement] Raise warning when dealing with numbers susceptible to precision errors

[YajJackson]

Describe the workflow you want to enable

You have some time series data that you wish to run through dbscan, only to find that your eps value is much lower than what produces expected results.

Consider

(Pdb) df['ts']
0     1.654272e+09
1     1.654272e+09
2     1.654272e+09
3     1.654272e+09
4     1.654272e+09
5     1.654272e+09
6     1.654272e+09
7     1.654272e+09
8     1.654272e+09
9     1.654272e+09
10    1.654272e+09
11    1.654272e+09
12    1.654272e+09
13    1.654272e+09
14    1.654272e+09
15    1.654272e+09
16    1.654272e+09
17    1.654272e+09
18    1.654272e+09
19    1.654272e+09
20    1.654272e+09
21    1.654272e+09
22    1.654272e+09
23    1.654272e+09
24    1.654273e+09
25    1.654273e+09
Name: ts, dtype: float64

(Pdb) max(np.diff(df['ts']))
8.0

It is reasonable for me to assume 8.0 as a fair eps value.

However

(Pdb) DBSCAN(eps=8, min_samples=10).fit_predict(df['ts'].to_numpy().reshape(-1,1))
array([-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
       -1, -1, -1, -1, -1, -1, -1, -1, -1])

Maybe you run some optimization that finds a value that works...

(Pdb) DBSCAN(eps=37, min_samples=10).fit_predict(df['ts'].to_numpy().reshape(-1,1))
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
       0, 0, 0, 0])

But 37 isn't a good value.

One suggestion is to proportionally scale your data/eps (which is totally fine! There are probably other ways the user should deal with this.)

(Pdb) DBSCAN(eps=8, min_samples=10).fit_predict(df['ts'].to_numpy().reshape(-1,1))
array([-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
       -1, -1, -1, -1, -1, -1, -1, -1, -1])
(Pdb) df['ts'] /= 1000
(Pdb) DBSCAN(eps=8 / 1000, min_samples=10).fit_predict(df['ts'].to_numpy().reshape(-1,1))
array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
       0, 0, 0, 0])

Describe your proposed solution

Warn users when they are performing potentially imprecise clustering.

Describe alternatives you've considered, if relevant

I recognize this as a fair user error that may be out of scope for the project, but allowing for some kind of raise_precision_warnings option might be very helpful to many users. Perhaps for other algorithms as well.

Additional context

• Numerical precision of euclidean_distances with float32 #9354
• DBSCAN not functioning properly with timestamp/large numbers #13006
• dbscan float precision #13604

[YajJackson]

After revisiting this, I found a preprocessing error on my end that was causing unexpected results. I'm going to close this because I am no longer able to reproduce the precision error I thought I was dealing with.