mi2graph

Generate Maximal Information Coefficient (MIC) relations between features with dynamic feature filtering for graph initialization.

Installation · Usage · Algorithm

---

Overview

mi2graph is a high-performance Rust tool for computing Maximal Information Coefficient (MIC) relationships between features in large-scale datasets. It implements dynamic feature filtering using sliding windows to identify optimal correlations and mutual information relationships.

Key Features

Feature	Description
Dynamic Sliding Windows	Adaptive window sizing for optimal correlation detection
Feature Filtering	Automatic removal of low-variation and redundant features
High Performance	Parallel processing with Rayon for multi-threaded execution
Memory Efficient	Streaming processing for large datasets
Standard Formats	Input/output in Apache Parquet and NumPy NPZ formats

Use Cases

• Bioinformatics: Gene expression network construction
• Machine Learning: Feature selection and graph neural network initialization
• Data Mining: Discovering non-linear relationships between variables

mi2graph is used for data preprocessing in our DeepTAN project.

---

Installation

Prerequisites

• Rust 1.70 or higher

From Crates.io

cargo install mi2graph

From Source

git clone https://github.com/WuChenh/mi2graph.git
cd mi2graph
cargo build --release

The compiled binary will be available at target/release/mi2graph.

---

Usage

Command Line Interface

mi2graph -i input.parquet -o output [OPTIONS]

Quick Start

# Basic usage
mi2graph -i data.parquet -o results

# With all options
mi2graph \
  --input data.parquet \
  --output results \
  --nfeat 5000 \
  --threcv 0.1 \
  --threpcc 0.95 \
  --thremi 0.05 \
  --chksim \
  --threads 8

Options

Option	Short	Default	Description
--input	-i	required	Input Parquet file path
--output	-o	required	Output path prefix
--nfeat	-n	5000	Number of top-CV features to retain (0 = use --threcv instead)
--threcv		0.1	Coefficient of variation threshold for feature filtering
--threpcc		0.95	Pearson correlation threshold for removing redundant features
--thremi		0.05	Mutual information threshold for edge filtering
--maxwin		0.98	Maximum window size ratio (window_size / num_samples)
--minwin		0.33	Minimum window size ratio
--stepwin		0.07	Window size step ratio
--stepsli		0.07	Sliding step ratio
--chksim	-s	false	Enable detection of similar feature pairs
--threads	-t	0	Number of threads (0 = use all available - 1)

---

Algorithm

Steps Explained

1. Coefficient of Variation (CV) Filtering

• Computes CV using dynamic sliding windows
• Removes features with low variation (noise)
• Can use either threshold-based or top-N selection

2. Similar Feature Detection (Optional)

• Uses Pearson correlation with dynamic 2D sliding windows
• Identifies and removes redundant features (highly correlated pairs)

3. Mutual Information Computation

• Computes MIC for all feature pairs using Freedman-Diaconis binning
• Uses dynamic sliding windows to find optimal mutual information
• Results are normalized (NMIC)

4. Edge Filtering

• Removes edges with MI below threshold
• Keeps only features involved in significant relationships

---

Input/Output Format

Input

The input must be a Parquet file with the following structure:

• Shape: (n_observations, 1 + n_features)
• First column: obs_names - observation/sample identifiers
• Remaining columns: feature values (numeric, f64)

obs_names	feature_1	feature_2	...	feature_n
sample_1	0.5	1.2	...	3.4
sample_2	0.7	0.9	...	2.1
...	...	...	...	...

Output

The tool generates two files:

1. NPZ File (output.npz)

Array Name	Shape	Description
mi_values	(n_edges,)	Sorted mutual information values
feat_pairs	(n_edges, 2)	Feature pair indices (edges)
processed_mat	(n_features, n_obs)	Filtered and processed data matrix
mat_feat_indices	(n_features,)	Original indices of retained features
mat_simi_feat_pairs	(n_similar_pairs, 2)	Indices of similar feature pairs
thre_cv, thre_pcc, thre_mi	(1,)	Threshold parameters used
ratio_*	(1,)	Window configuration parameters

2. Parquet File (output.parquet)

A DataFrame containing the processed matrix with observation names and feature names.

---

Library Usage

You can also use mi2graph as a Rust library:

use mi2graph::{
    mic_mat_with_data_filter, read_parquet_to_array2d,
    FilterConfig, ProcessingOptions, WindowConfig
};

fn main() -> Result<(), Box<dyn std::error::Error>> {
    // Read data
    let (data, obs_names, var_names) = read_parquet_to_array2d("input.parquet")?;
    
    // Configure processing
    let filter_config = FilterConfig {
        thre_cv: 0.1,
        thre_pcc: 0.95,
        thre_mi: 0.05,
    };
    
    let window_config = WindowConfig {
        ratio_min: 0.33,
        ratio_max: 0.98,
        ratio_step: 0.07,
        ratio_slide: 0.07,
    };
    
    let options = ProcessingOptions {
        check_sim: true,
        n_features_to_select: 5000,
        n_threads: 8,
    };
    
    // Run analysis
    mic_mat_with_data_filter(
        "output",
        &data,
        &obs_names,
        &var_names,
        filter_config,
        window_config,
        options,
    )?;
    
    Ok(())
}

Add to your Cargo.toml:

[dependencies]
mi2graph = "0.3"

---

License

This project is licensed under either of:

• MIT License - See LICENSE file
• Apache License, Version 2.0 - See LICENSE file

at your option.