OPTICS

Ordering Points To Identify Clustering Structure - density-based algorithm similar to DBSCAN but doesn't require preset eps parameter.

When to use:

• Have clusters of varying densities
• Don't want to tune eps parameter
• Need hierarchical view of density-based clusters
• Want more robust than DBSCAN

Strengths: Handles varying densities, more robust than DBSCAN, creates reachability plot, automatic parameter selection Weaknesses: Slower than DBSCAN, more complex to interpret, requires more memory

Model Parameters

Min Samples (default: 5) Minimum points in a neighborhood to be considered a core point.

• 3-5: Sensitive to local structure
• 5-10: Good default
• 10+: More conservative, larger clusters

Max Eps (optional) Maximum distance between two samples for one to be considered a neighbor.

• null: No limit (examines all distances)
• Set value: Limits neighborhood size for speed

Metric (default: "minkowski") Distance metric:

• minkowski: Generalized distance (default)
• euclidean: Standard distance
• manhattan: City-block distance
• chebyshev: Maximum coordinate difference

Cluster Method (default: "xi") How to extract clusters from reachability plot:

• xi: Automatic extraction using steepness (default, better)
• dbscan: Extract using eps threshold (similar to DBSCAN)

Algorithm (default: "auto") Nearest neighbor algorithm:

• auto: Automatically choose best
• ball_tree: Good for low-medium dimensions
• kd_tree: Fast for low dimensions
• brute: Exact but slow