dyf 0.1.3

Density Yields Features - discover structure in embedding spaces

DYF - Outlier Classification

Fast outlier classification using PCA-based LSH. Identifies three types of items in embedding spaces:

• Dense: Items in well-populated semantic buckets (the majority)
• Diaspora: Sparse items that find community via recovery PCA (misplaced by global structure)
• Orphan: Truly unique items with no semantic neighbors

Installation

pip install dyf

For full features (embedding generation, LLM labeling):

pip install dyf[full]

Quick Start

Fast Classification (Rust-accelerated)

import numpy as np
from dyf import OutlierClassifier

# Create embeddings (e.g., from sentence-transformers)
embeddings = np.random.randn(10000, 384).astype(np.float32)

# Classify outliers (~60ms for 60K samples)
classifier = OutlierClassifier(embedding_dim=384)
classifier.fit(embeddings)

# Get results
print(classifier.report())
diaspora = classifier.get_diaspora()  # Indices of diaspora items
orphans = classifier.get_orphans()    # Indices of orphan items

Full-Featured Usage (with embeddings & labeling)

from dyf import OutlierClassifierFull, EmbedderConfig, LabelerConfig

# From raw texts with built-in TF-IDF embeddings
classifier = OutlierClassifierFull.from_texts(
    texts=documents,
    categories=categories,  # Optional category labels
    embedding_dim=128
)

# Or use sentence-transformers
embeddings = EmbedderConfig.MEDIUM.embed(texts)  # all-mpnet-base-v2
classifier = OutlierClassifierFull(embedding_dim=768)
classifier.fit(embeddings, categories=categories, texts=texts)

# Get detailed report
print(classifier.report())

# Label buckets with local LLM
labels = classifier.label_buckets(**LabelerConfig.MEDIUM.as_kwargs())
print(labels['dense'][1234]['label'])  # "Reinforcement Learning"

# Or use keyword extraction (no LLM required)
labels = classifier.label_buckets_keywords()

Performance

Implementation	60K samples (384d)	Per sample
DYF (Rust)	~60ms	1.0 µs
Pure Python	~230ms	3.8 µs

3.8x faster than pure Python/sklearn.

API Reference

OutlierClassifier (Fast)

OutlierClassifier(
    embedding_dim: int,
    initial_bits: int = 14,       # Bits for initial PCA LSH
    recovery_bits: int = 8,       # Bits for recovery PCA
    dense_threshold: int = 10,    # Min bucket size for "dense"
    intra_outlier_std: float = 2.0,   # Std threshold for intra-bucket outliers
    recovery_cluster_min: int = 3,    # Min cluster size for "recovered"
    seed: int = 31
)

Methods:

• fit(embeddings) - Fit on numpy array (n_samples, embedding_dim)
• fit_arrow(arrow_array) - Fit on PyArrow FixedSizeListArray (zero-copy)
• get_diaspora() - Get indices of diaspora items
• get_orphans() - Get indices of orphan items
• get_statuses() - Get status for all items
• report() - Get classification report

EmbedderConfig Presets

Name	Model	Dimensions	Size
TFIDF	TF-IDF + SVD	128	0 MB
LOW	all-MiniLM-L6-v2	384	80 MB
MEDIUM	all-mpnet-base-v2	768	420 MB
HIGH	bge-large-en-v1.5	1024	1.3 GB

LabelerConfig Presets

Name	Model	Parameters
KEYWORDS	TF-IDF keywords	-
LOW	phi3:mini	3.8B
MEDIUM	qwen2.5:7b	7B
HIGH	qwen2.5:14b	14B

Algorithm

Two-stage PCA-based LSH outlier classification:

1. Stage 1: Random hash → bucket centroids → PCA on centroids → re-hash
2. Outlier Detection: Sparse buckets + intra-bucket distance outliers
3. Stage 2: Recovery PCA on outliers → diaspora vs orphan

The key insight: outliers from global PCA often share structure at coarser resolution.

License

MIT