Available Partitioning Methods

The SHiP framework provides multiple partitioning methods to extract flat clusterings from hierarchies. Each method corresponds to a specific partitioning objective or heuristic, such as fixed-\(k\), elbow detection, and also function-based selection strategies, as e.g., the HDBSCAN stability function.

The enum below shows all available partitioning strategies:

enum class PartitioningMethod

Values:

enumerator K

Partition into a fixed number of clusters \(k\).

**Parameters:**
- `k` (default: `2`): Number of clusters the data should be partitioned into.

enumerator Elbow

Use the Elbow method to determine the optimal number of clusters.

enumerator Threshold

Cut the hierarchy at a fixed similarity threshold.

**Parameters:**
- `k` (default: `2`): Number of clusters the data should be partitioned into.

enumerator ThresholdElbow

Combine threshold cutting and the Elbow method for more adaptive partitioning.

enumerator QCoverage

Use \(q\)-coverage to ensure that a given proportion of mass is retained in the clustering.

**Parameters:**
- `k` (default: `2`): Number of clusters the data should be partitioned into.
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.

enumerator QCoverageElbow

Combine \(q\)-coverage and the Elbow method for adaptive and coverage-aware partitioning.

**Parameters:**
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.

enumerator QStem

Use the \(q\)-stem criterion to extract partitions based on internal branch structure.

**Parameters:**
- `k` (default: `2`): Number of clusters the data should be partitioned into.
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.

enumerator QStemElbow

Combine \(q\)-stem with Elbow-based refinement for improved robustness.

**Parameters:**
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.

enumerator LcaNoiseElbow

Elbow method that incorporates noise suppression using least common ancestors (LCA).

enumerator LcaNoiseElbowNoTriangle

Noise-aware Elbow method that excludes triangular merging to improve sharpness.

enumerator MedianOfElbows

Take the median of multiple Elbow criteria to find a more stable clustering.

**Parameters:**
- `elbow_start_z` (default: `1`): Range start to use for elbow method.
- `elbow_end_z` (default: `5`): Range end to use for elbow method.

enumerator MeanOfElbows

Take the mean of multiple Elbow criteria to smooth over noise in the hierarchy.

**Parameters:**
- `elbow_start_z` (default: `1`): Range start to use for elbow method.
- `elbow_end_z` (default: `5`): Range end to use for elbow method.

enumerator Stability

Optimize cluster stability across multiple resolutions to find a robust partitioning, see also HDBSCAN.

**Parameters:**
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.

enumerator NormalizedStability

Use normalized cluster stability to allow comparisons across hierarchies of different sizes.

**Parameters:**
- `min_cluster_size` (default: `5`): Minimum number of points which a cluster should contain.