Introduction & Problem Setup

• Presenter: Josh Starmer from StatQuest
• Topic: DBSCAN clustering algorithm explanation
• Example dataset: weight (x-axis) and height (y-axis) measurements from people
• Visual identification: 2 distinct clusters visible by eye
  • Dense clumps of people forming clusters
  • Outliers: people far from everyone else

Limitations of Traditional Clustering Methods

• Problem: nested clusters (green cluster wraps around blue cluster)
• K-means clustering struggles with nested structures
  • May assign points incorrectly due to nesting
  • Points that visually belong to green cluster assigned to blue
• High-dimensional challenge:
  • 2D: weight + height (drawable)
  • 3D: weight + height + age (awkward but possible)
  • 4D+: impossible to visualize, need algorithmic solution

DBSCAN Algorithm Overview

• Purpose: identify nested clusters in high dimensions
• Mimics human visual clustering ability
• Core principle: uses point density to identify clusters
  • High density regions = clusters
  • Low density regions = outliers

DBSCAN Step-by-Step Process

Step 1: Count Nearby Points

• Draw circle around each point (user-defined radius parameter)
• Count points that overlap with circle (at least partially)
• Examples from demonstration:
  • Red point 1: close to 8 other points
  • Red point 2: close to 5 other points
  • Red point 3: close to 6 other points
  • Red point 4: close to 7 other points
  • Red point 5: close to 2 other points
  • Red point 6: close to 0 other points

Step 2: Define Core Points

• Core point definition: close to at least 4 other points (user-defined parameter)
• Classification results:
  • Core points: points with ≥4 neighbors
  • Non-core points: points with <4 neighbors

Step 3: Build First Cluster

• Randomly select core point → assign to first cluster
• Add neighboring core points that overlap with orange circle
• Core points extend cluster to other nearby core points
• Process continues until no more core points can be added
• Result: all points in first cluster are core points

Step 4: Add Non-Core Points to First Cluster

• Add non-core points close to core points in first cluster
• Important rule: non-core points can only join clusters, cannot extend them
• Non-core points not used to find additional cluster members
• Example: non-core point close to another non-core point (already in cluster) will NOT be added if not close to core point

Step 5: Create Additional Clusters

• Remaining unassigned core points form new clusters
• Core points close to each other grouped together
• Non-core points close to new cluster core points added
• Process repeats until all core points assigned

Step 6: Handle Outliers

• Remaining non-core points not close to any core points = outliers
• Not assigned to any cluster

Important Algorithm Properties

• Sequential cluster creation
• Point assignment priority: if non-core point close to multiple clusters, assigned to first cluster built
• Once assigned, points cannot move to other clusters

Key Parameters (User-Defined)

• Circle radius for determining "closeness"
• Minimum number of points for core point definition (example used: 4)
• May require parameter adjustment/experimentation

Self-Promotion Segment

• StatQuest study guides available at statquest.org
• Support options: Patreon, channel membership, merchandise, donations