DBSCAN clustering in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - DBSCAN clustering

Which metric matters for DBSCAN clustering and WHY

DBSCAN is a clustering method that groups points based on density. It finds clusters of points close together and marks points in low-density areas as noise.

Since DBSCAN is unsupervised (no true labels), common metrics like accuracy don't apply. Instead, we use metrics that compare the clustering result to known labels if available, or measure cluster quality internally.

Key metrics include:

• Silhouette Score: Measures how well points fit their cluster compared to others. Values near 1 mean good clusters, near 0 means overlapping clusters, and negative means wrong clusters.
• Adjusted Rand Index (ARI): Compares clustering to true labels (if known). It adjusts for chance and ranges from -1 to 1, where 1 is perfect match.
• Number of clusters and noise points: Helps understand if DBSCAN found meaningful groups or too many/few clusters.

These metrics help us know if DBSCAN found meaningful groups or just noise.

Confusion matrix or equivalent visualization

DBSCAN does not produce a confusion matrix directly because it is unsupervised. However, if true labels are known, we can compare clusters to labels using a contingency table.

    Contingency Table Example:

    | Cluster \ Label | Label A | Label B | Label C |
    |-----------------|---------|---------|---------|
    | Cluster 1       |   30    |    2    |    1    |
    | Cluster 2       |    3    |   25    |    0    |
    | Noise (-1)      |    2    |    3    |   10    |

    This table shows how many points from each true label fall into each cluster or noise.
    

From this, metrics like Adjusted Rand Index or Normalized Mutual Information can be calculated.

Precision vs Recall tradeoff with concrete examples

In clustering, precision and recall relate to how well clusters capture true groups.

For DBSCAN:

• High precision: Points in a cluster mostly belong to the same true group. Few wrong points included.
• High recall: Most points of a true group are found in the cluster. Few points missed or marked as noise.

Example:

• If DBSCAN is strict (small radius), it may create many small clusters and label many points as noise. This gives high precision (clusters are pure) but low recall (many true points missed).
• If DBSCAN is loose (large radius), it may merge groups and include noise points. This gives high recall (most points grouped) but low precision (clusters mix different groups).

Choosing parameters balances this tradeoff depending on what matters more: pure clusters or complete groups.

What "good" vs "bad" metric values look like for DBSCAN clustering

Good DBSCAN clustering:

• Silhouette score close to 0.5 or higher (up to 1). This means clusters are well separated and dense.
• Adjusted Rand Index near 1 if true labels exist, showing clusters match real groups.
• Reasonable number of clusters (not too many or too few) and low noise percentage.

Bad DBSCAN clustering:

• Silhouette score near 0 or negative, meaning clusters overlap or are poorly formed.
• Adjusted Rand Index near 0 or negative, meaning clusters do not match true groups.
• Too many clusters (over-segmentation) or too few clusters (under-segmentation), or very high noise rate.

Common pitfalls in DBSCAN metrics

• Ignoring noise points: Noise points are not assigned to clusters but affect metrics. Counting or ignoring them changes results.
• Parameter sensitivity: DBSCAN depends on radius and minimum points. Poor choices lead to misleading metrics.
• Using accuracy: Accuracy is not meaningful for clustering because labels are unknown or arbitrary.
• Overfitting clusters: Too many small clusters can inflate some metrics but do not represent meaningful groups.
• Data scale and density: DBSCAN assumes similar density clusters. Varying densities cause poor clustering and metrics.

Self-check question

Your DBSCAN model found 5 clusters and labeled 30% of points as noise. The silhouette score is 0.6. Is this good?

Answer: A silhouette score of 0.6 is quite good, showing well-formed clusters. Having 30% noise can be reasonable if the data has outliers or sparse areas. The number of clusters should be checked if it matches expected groups. Overall, this suggests DBSCAN did a good job, but review parameters and domain knowledge to confirm.

Key Result

Silhouette score and Adjusted Rand Index are key metrics to evaluate DBSCAN clustering quality and match to true groups.

Practice

1. What is the main advantage of using DBSCAN clustering compared to K-Means?

easy

A. It does not require specifying the number of clusters beforehand.

B. It always produces clusters of equal size.

C. It only works with numerical data.

D. It requires labeled data to work.

Solution

1. Step 1: Understand DBSCAN's clustering approach

   DBSCAN groups points based on density and does not need the number of clusters as input.

2. Step 2: Compare with K-Means

   K-Means requires the number of clusters to be specified before running the algorithm.

3. Final Answer:

   It does not require specifying the number of clusters beforehand. -> Option A

4. Quick Check:

   DBSCAN = no cluster count needed [OK]

Hint: DBSCAN finds clusters by density, no cluster count needed [OK]

Common Mistakes:

• Thinking DBSCAN needs number of clusters like K-Means
• Assuming DBSCAN only works with numerical data
• Believing DBSCAN requires labeled data

2. Which of the following is the correct way to import and create a DBSCAN model in Python using scikit-learn?

easy

A. import DBSCAN from sklearn\ndbscan = DBSCAN(0.5, 5)

B. import dbscan from sklearn.cluster\ndbscan = dbscan(eps=0.5, min_samples=5)

C. from sklearn.cluster import dbscan\ndbscan = dbscan(eps=0.5, min_samples=5)

D. from sklearn.cluster import DBSCAN\ndbscan = DBSCAN(eps=0.5, min_samples=5)

Solution

1. Step 1: Check correct import syntax

   The correct import is from sklearn.cluster import DBSCAN with capital letters.

2. Step 2: Check model creation syntax

   DBSCAN is created by calling DBSCAN(eps=0.5, min_samples=5) with named parameters.

3. Final Answer:

   from sklearn.cluster import DBSCAN\ndbscan = DBSCAN(eps=0.5, min_samples=5) -> Option D

4. Quick Check:

   Correct import and parameters = from sklearn.cluster import DBSCAN\ndbscan = DBSCAN(eps=0.5, min_samples=5) [OK]

Hint: Use capital DBSCAN and named parameters eps, min_samples [OK]

Common Mistakes:

• Using lowercase dbscan instead of DBSCAN
• Wrong import syntax like import dbscan from sklearn.cluster
• Passing parameters without names or in wrong order

3. Given the following code, what will be the output of print(labels)?

from sklearn.cluster import DBSCAN
import numpy as np

X = np.array([[1, 2], [2, 2], [8, 7], [8, 8], [25, 80]])
dbscan = DBSCAN(eps=3, min_samples=2)
dbscan.fit(X)
labels = dbscan.labels_
print(labels)

medium

A. [0 0 1 1 -1]

B. [0 1 0 1 -1]

C. [1 1 0 0 -1]

D. [-1 -1 -1 -1 -1]

Solution

1. Step 1: Understand DBSCAN parameters and data points

   eps=3 means points within distance 3 are neighbors; min_samples=2 means at least 2 points to form a cluster. Points [1,2] and [2,2] are close, so cluster 0. Points [8,7] and [8,8] are close, so cluster 1. Point [25,80] is far and isolated, so noise (-1).

2. Step 2: Assign labels based on clusters and noise

   First cluster points get label 0, second cluster 1, noise points get -1.

3. Final Answer:

   [0 0 1 1 -1] -> Option A

4. Quick Check:

   Clusters and noise labels = [0 0 1 1 -1] [OK]

Hint: Close points cluster, isolated points labeled -1 noise [OK]

Common Mistakes:

• Mixing cluster labels order
• Forgetting noise points get label -1
• Assuming all points form one cluster

4. You run DBSCAN with eps=0.5 and min_samples=3 on a dataset but get all points labeled as -1. What is the most likely cause?

medium

A. The dataset has too many clusters.

B. The eps value is too small, so no points have enough neighbors.

C. The min_samples value is too low, causing all points to be noise.

D. DBSCAN requires the number of clusters to be set manually.

Solution

1. Step 1: Analyze effect of eps and min_samples

   eps controls neighborhood radius; if too small, points have few neighbors. min_samples=3 means at least 3 points needed to form a cluster.

2. Step 2: Reason why all points are noise

   If eps is too small, no point has 3 neighbors within that radius, so all points become noise (-1).

3. Final Answer:

   The eps value is too small, so no points have enough neighbors. -> Option B

4. Quick Check:

   Small eps causes all noise = The eps value is too small, so no points have enough neighbors. [OK]

Hint: Too small eps means no clusters, all points noise [OK]

Common Mistakes:

• Thinking min_samples too low causes noise
• Believing DBSCAN needs cluster count input
• Assuming many clusters cause noise labels

5. You have a dataset with two dense clusters and some scattered noise points. You want DBSCAN to find the clusters and mark noise correctly. Which combination of eps and min_samples is best to try first?

hard

A. eps=0.3, min_samples=10

B. eps=5, min_samples=2

C. eps=1, min_samples=5

D. eps=0.1, min_samples=1

Solution

1. Step 1: Understand parameter effects

   eps controls neighborhood size; min_samples controls minimum points to form cluster. Too small eps or too high min_samples may miss clusters; too large eps or too low min_samples may merge clusters or label noise incorrectly.

2. Step 2: Choose balanced parameters

   eps=1 and min_samples=5 is a common starting point balancing cluster detection and noise filtering for dense clusters.

3. Final Answer:

   eps=1, min_samples=5 -> Option C

4. Quick Check:

   Balanced eps and min_samples = eps=1, min_samples=5 [OK]

Hint: Start with moderate eps=1 and min_samples=5 for dense clusters [OK]

Common Mistakes:

• Using too small eps causing all noise
• Using too low min_samples merging noise as clusters
• Choosing very large eps merging clusters incorrectly