PaCMAP<T>
Models & Types ยท AiDotNet.Preprocessing.DimensionalityReduction
PaCMAP: Pairwise Controlled Manifold Approximation.
Inheritance: TransformerBase<T, Matrix<T>, Matrix<T>> โ PaCMAP<T>
For Beginners
PaCMAP improves on UMAP/t-SNE by:
- Preserving global structure better through mid-near and further pairs
- Using controlled pair selection instead of random sampling
- Dynamically adjusting focus from global to local structure
- Being more robust to hyperparameter choices
Use cases:
- When you need faithful global structure
- When t-SNE/UMAP produces fragmented clusters
- When relative distances between clusters matter
- Biological data, image embeddings, document visualization
How It Works
PaCMAP is a dimensionality reduction method designed to preserve both local and global structure. It uses three types of point pairs with carefully controlled weights during optimization to achieve better structure preservation than t-SNE or UMAP.
The algorithm:
- Create three types of pairs: nearby (local), mid-near (intermediate), further (global)
- Initialize embedding using PCA
- Optimize using attractive/repulsive forces with dynamic weighting
- Gradually shift from global to local focus during optimization
Constructors
| Constructor | Summary |
|---|---|
PaCMAP(int nComponents, int nNeighbors, double mnRatio, double fpRatio, int nIter, double learningRate, int? randomState, int[] columnIndices) |
Creates a new instance of PaCMAP. |
Properties
| Property | Summary |
|---|---|
Matrix<T> Embedding { get; } |
Gets the embedding result. |
bool SupportsInverseTransform { get; } |
Gets whether this transformer supports inverse transformation. |
int NComponents { get; } |
Gets the number of components (dimensions). |
int NNeighbors { get; } |
Gets the number of neighbors. |
Methods
| Method | Summary |
|---|---|
string[] GetFeatureNamesOut(string[] inputFeatureNames) |
Gets the output feature names after transformation. |