Course contentsShow
← Machine Learning and Deep Learning
Lesson 350 of 3,5388. Unsupervised Learning: ClusteringFree lesson

Choosing Epsilon and MinPts Parameters

Strategies for selecting DBSCAN hyperparameters using k-distance plots and domain knowledge.

Choosing Epsilon and MinPts Parameters

What you'll learn: How to select the right epsilon (ε) and MinPts hyperparameters for DBSCAN using practical strategies.

The Challenge

DBSCAN's performance hinges on two critical parameters: epsilon (the neighborhood radius) and MinPts (minimum points to form a core point). Unlike K-Means where you pick K clusters, here you're defining what "dense enough" means—and choosing poorly can fragment natural clusters or merge distinct ones.

Strategy 1: The K-Distance Plot

The most reliable approach uses a k-distance graph:

  1. For each point, calculate the distance to its k-th nearest neighbor (where k = MinPts)
  2. Sort these distances in ascending order
  3. Plot them as a curve
  4. Look for the "elbow"—where the curve sharply rises

The elbow indicates where distances jump significantly, suggesting the boundary between dense regions and noise. This inflection point gives you a good epsilon value.

Think of it like: Imagine plotting home prices sorted by distance from downtown. The elbow shows where "downtown neighborhoods" end and "suburbs" begin—that's your natural boundary.

Strategy 2: Domain Knowledge

Use what you know about your problem:

  • MinPts: A common rule of thumb is MinPts = 2 × dimensions, with a minimum of 4 for 2D data
  • Epsilon: Consider the physical meaning—if clustering GPS coordinates, what distance (in meters) defines "nearby"?

Practical Tips

  • Start with the k-distance plot for epsilon, then fine-tune
  • Higher dimensions need larger MinPts to be statistically meaningful
  • Test multiple values and compare using Silhouette Score

Key Takeaway: Use k-distance plots to find the "elbow" for epsilon selection, and apply domain knowledge for MinPts—typically starting with 2× your feature dimensions.