SNIC: Sparse Network of Idempotent Comparisons

A high-performance Python library for efficient ranking of large datasets using sparse comparison networks.

What is SNIC?

SNIC is a novel ranking algorithm that dramatically reduces the number of comparisons needed to rank large sets of elements. Instead of requiring O(n²) pairwise comparisons like traditional methods, SNIC achieves O(n log n) efficiency through:

• Sparse Networks: Structured comparison networks that capture ranking information with minimal comparisons
• Base-sized Matchups: Fixed-size groups that maximize information extraction per comparison
• GBER Foundation: Mathematical framework using Generalized Base Exponential Representation for optimal network structure

Key Benefits

• Scalable: Efficiently handles datasets from hundreds to millions of elements
• Sparse: Requires far fewer comparisons than traditional ranking methods
• Structured: Deterministic, mathematically-grounded approach to comparison generation
• Fast: High-performance Rust implementation with Python bindings

Quick Start

pip install snic

import snic

# Generate matchups for ranking 100 elements with base-3 comparisons
matchups = snic.stream_matches_from(100, 3)

# Each matchup contains 3 elements to be ranked
print(f"Generated {len(matchups)} matchups")
print(f"First matchup: {matchups[0]}")  # e.g., [0, 33, 67]

# After ranking each matchup externally (by human judgment, ML model, etc.)
# Convert ranked matchups back to final ranking
ranked_matchups = [
    [1, 0, 2],  # Example: element 1 ranked first, 0 second, 2 third
    [4, 3, 5],  # Continue for all matchups...
    # ... 
]

final_ranking = snic.stream_rankings_from(ranked_matchups)
print(f"Final ranking: {final_ranking}")

How It Works

1. Decomposition: SNIC uses GBER to break down your dataset size into optimal subnetworks
2. Matchup Generation: Creates structured groups of elements for comparison
3. Sparse Comparisons: Each element participates in logarithmically few comparisons
4. Ranking Synthesis: Combines local rankings into a global result

Use Cases

• Large-scale surveys: Rank thousands of items with minimal human effort
• Content recommendation: Efficiently determine user preferences
• Tournament systems: Fair bracket generation for competitions
• Data analysis: Rank features, samples, or model outputs at scale

Development

Local Setup

# Clone the repository
git clone https://github.com/ryzhakar/snic-rs
cd snic-rs

# Install development dependencies
pip install maturin

# Build and install in development mode
maturin develop

# Run tests
python -m pytest

Building

# Build wheel
maturin build --release

# Build and publish to PyPI
maturin publish

Algorithm Details

SNIC is based on rigorous mathematical foundations:

• GBER (Generalized Base Exponential Representation): Decomposes dataset sizes into optimal subnetwork structures
• Stride-based Matchup Generation: Ensures comprehensive coverage with minimal redundancy
• Hub-and-Spoke Connectivity: Connects subnetworks for global ranking consistency

For detailed algorithm documentation, see the wiki.

Performance

SNIC's efficiency scales logarithmically:

Dataset Size	Traditional Comparisons	SNIC Comparisons	Reduction
1,000	499,500	~3,000	99.4%
10,000	49,995,000	~40,000	99.9%
100,000	4,999,950,000	~500,000	99.99%

License

AGPLv3 License - see LICENSE for details.