Not every point in a dataset has to belong to a cluster. DBSCAN can group dense regions while labelling some observations as noise, which makes it useful to understand when a method that assigns every point feels too rigid.
The scikit-learn DBSCAN documentation describes two central controls: eps, which sets the neighbourhood distance, and min_samples, which sets the support needed for a core point. The latter includes the point itself in scikit-learn’s definition. The algorithm does not ask you to specify a cluster count in advance.
A neighbourhood is not a maximum cluster diameter
Imagine a curved chain of closely spaced observations. Neighbourhood connections can link the dense parts into one cluster even though its endpoints are farther apart than eps. Interpreting that parameter as the maximum distance between every pair of members would give the wrong mental model.
Points can be core points, border points reached from a dense region, or noise under the chosen settings. “Noise” is an algorithmic label for this run, not proof that a record is erroneous or unimportant. A rare but valid event can be exactly the observation a project needs to examine.
Inspect the assumptions behind distance
Choose features and a distance measure that make sense for the task. Units and scaling affect neighbourhoods. A radius that is meaningful for geographical coordinates in one representation may be meaningless for a mixed table of ages, prices and category codes.
Compare a few justified parameter settings and inspect what changes. If one setting merges several visibly distinct regions while another leaves almost everything as noise, record that sensitivity. Uneven density and high-dimensional data can make a single neighbourhood scale difficult to choose.
Finally, explain the labels in the vocabulary of the application and test whether they support a useful decision. DBSCAN can reveal a pattern in the selected representation; it does not establish natural categories merely because it produces a colourful plot. Keep the feature preparation, metric and parameters with the result so another analyst can reproduce the same interpretation.