pyppann

Approximate Nearest Neighbors using Projection Pursuit.

pyppann provides ANN methods that use learned projections to find approximate nearest neighbors. The key innovation is NeighborPreservingANN, which learns projections that directly optimize neighbor preservation (recall@k) rather than variance or non-Gaussianity.

Installation

pip install pyppann

Quick Start

import numpy as np
from pyppann import NeighborPreservingANN

# Generate sample data
X = np.random.randn(1000, 50)

# Fit the model
model = NeighborPreservingANN(n_components=20, k=10)
model.fit(X, k=10)

# Find k nearest neighbors for queries
queries = np.random.randn(5, 50)
distances, indices = model.kneighbors(queries, n_neighbors=10)

Available Methods

Class	Description
NeighborPreservingANN	Learns projections that maximize neighbor preservation. Best recall.
ProjectionPursuitANN	Uses projection pursuit regression via pyppur.
ANNOYWrapper	Wrapper around the ANNOY library with sklearn-compatible API.
RandomProjectionANN	Gaussian random projections + brute kNN. Simple baseline.
PCAAnn	PCA dimensionality reduction + brute kNN. Simple baseline.

Lower-level API

NeighborPreservingPP provides a scikit-learn compatible transformer (fit/transform) for the neighbor-preserving projection pursuit algorithm:

from pyppann import NeighborPreservingPP

pp = NeighborPreservingPP(n_components=20, k=10)
X_projected = pp.fit_transform(X)

Evaluation Utilities

pyppann includes utilities for benchmarking ANN methods:

from pyppann import compute_recall, run_comparison, run_full_evaluation

# Compare methods on synthetic data
results = run_comparison(n_components=20, k=10, verbose=True)

# Full evaluation across multiple k and n_components values
all_results = run_full_evaluation(
    n_samples=5000,
    n_features=50,
    k_values=(5, 10, 20),
    n_component_values=(10, 20, 30),
)

License

MIT License