2 unstable releases

0.2.0 Jun 24, 2022
0.1.1 Feb 12, 2022

#827 in Algorithms

GPL-3.0 license

1MB
1K SLoC

Rust affinityprop crate affinityprop docs GitHub affinityprop: rustc 1.58 coverage

affinityprop

The affinityprop crate provides an optimized implementation of the Affinity Propagation clustering algorithm, which identifies cluster of data without a priori knowledge about the number of clusters in the data. The original algorithm was developed by Brendan Frey and Delbery Dueck

About

Affinity Propagation identifies a subset of representative examples from a dataset, known as exemplars.

Briefly, the algorithm accepts as input a matrix describing pairwise similarity for all data values. This information is used to calculate pairwise responsibility and availability. Responsibility r(i,j) describes how well-suited point j is to act as an exemplar for point i when compared to other potential exemplars. Availability a(i,j) describes how appropriate it is for point i to accept point j as its exemplar when compared to other exemplars.

Users provide a number of convergence iterations to repeat the calculations, after which the potential exemplars are extracted from the dataset. Then, the algorithm continues to repeat until the exemplar values stop changing, or until the maximum iterations are met.

Why this crate?

The nature of Affinity Propagation demands an O(n2) runtime. An existing sklearn version is implemented using the Python library numpy which incorporates vectorized row operations. Coupled with SIMD instructions, this results in decreased time to finish.

However, in applications with large input values, the O(n2) runtime is still prohibitive. This crate implements Affinity Propagation using the rayon crate, which allows for a drastic decrease in overall runtime - as much as 30-60% when compiled in release mode!

Dependencies

cargo with rustc >=1.58

Installation

In Rust code

[dependencies]
affinityprop = "0.2.0"
ndarray = "0.15.4"

As a command-line tool

cargo install affinityprop

Usage

From Rust code

The affinityprop crate expects a type that defines how to calculate pairwise Similarity for all data points. This crate provides the NegEuclidean, NegCosine, and LogEuclidean structs, which are defined as -1 * sum((a - b)**2), -1 * (a . b)/(|a|*|b|), and sum(log((a - b)**2)), respectively.

Users who wish to calculate similarity differently are advised that Affinity Propagation expects s(i,j) > s(i, k) iff i is more similar to j than it is to k.

use ndarray::{arr1, arr2, Array2};
use affinityprop::{AffinityPropagation, NegCosine, Preference};

let x: Array2<f32> = arr2(&[[0., 1., 0.], [2., 3., 2.], [3., 2., 3.]]);

// Cluster using negative cosine similarity with a pre-defined preference
let ap = AffinityPropagation::default();
let (converged, results) = ap.predict(&x, NegCosine::default(), Preference::Value(-10.));
assert!(converged && results.len() == 1 && results.contains_key(&0));

// Cluster with list of preference values
let pref = arr1(&[0., -1., 0.]);
let (converged, results) = ap.predict(&x, NegCosine::default(), Preference::List(&pref));
assert!(converged);
assert!(results.len() == 2 && results.contains_key(&0) && results.contains_key(&2));

// Use damping=0.5, threads=2, convergence_iter=10, max_iterations=100,
// median similarity as preference
let ap = AffinityPropagation::new(0.5, 2, 10, 100);
let (converged, results) = ap.predict(&x, NegCosine::default(), Preference::Median);
assert!(converged);
assert!(results.len() == 2 && results.contains_key(&0) && results.contains_key(&2));

// Predict with pre-calculated similarity
let s: Array2<f32> = arr2(&[[0., -3., -12.], [-3., 0., -3.], [-12., -3., 0.]]);
let ap = AffinityPropagation::default();
let (converged, results) = ap.predict_precalculated(s, Preference::Value(-10.));
assert!(converged && results.len() == 1 && results.contains_key(&1));

From the Command Line

affinityprop can be run from the command-line and used to analyze a file of data:

ID1 val1 val2
ID2 val3 val4
ID3 val5 val6

where IDn is any string identifier and valn are floating-point (decimal) values. The file delimiter is provided from the command line with the -l flag. Similarity will be calculated based on the option set by the -s flag.

For files without row ids:

val1 val2
val3 val4
val5 val6

provide the -n flag from the command line. IDs will automatically be assigned by zero-based index.

Users may instead provide a pre-calculated similarity matrix by passing the -s 3 flag from the command line and by structuring their input file as:

ID1 sim11 sim12 sim13
ID2 sim21 sim22 sim23
ID3 sim31 sim32 sim33

where rowi, colj is the pairwise similarity between inputs i and j.

Or, for files without row labels, users may pass -n -s 3:

sim11 sim12 sim13
sim21 sim22 sim23
sim31 sim32 sim33

IDs will automatically be assigned by zero-based index.

Help Menu

affinityprop 0.2.0
Chris N. <christopher.neely1200@gmail.com>
Vectorized and Parallelized Affinity Propagation

USAGE:
    affinityprop [OPTIONS] --input <INPUT>

FLAGS:
    -n, --no_labels    Input file does not contain IDS as the first column
    -h, --help         Prints help information
    -V, --version      Prints version information

OPTIONS:
    -c, --convergence_iter <CONV_ITER>    Convergence iterations, default=10
    -d, --damping <DAMPING>               Damping value in range (0, 1), default=0.9
    -l, --delimiter <DELIMITER>           File delimiter, default '\t'
    -i, --input <INPUT>                   Path to input file
    -m, --max_iter <MAX_ITER>             Maximum iterations, default=100
    -r, --precision <PRECISION>           Set f32 or f64 precision, default=f32
    -p, --preference <PREF>               Preference to be own exemplar, default=median pairwise similarity
    -s, --similarity <SIMILARITY>         Set similarity calculation method
                                          (0=NegEuclidean,1=NegCosine,2=LogEuclidean,3=precalculated), default=0
    -t, --threads <THREADS>               Number of worker threads, default=4

Results

Results are printed to stdout in the format:

Converged=true/false nClusters=NC nSamples=NS
>Cluster=n size=N exemplar=i
[comma-separated cluster member IDs/indices]
>Cluster=n size=N exemplar=i
[comma-separated cluster member IDs/indices]
...

Runtime and Resource Notes

Affinity Propagation is O(n2) in both runtime and memory. This crate seeks to address the former, not the latter.

An estimated memory usage can be calculated given:

memory(GB) = p * 4 * N^2 / 2^30

For N inputs. p = 4 for 32-bit floating-point precision and p = 8 for 64-bit.

Comparison to sklearn implementation

This implementation was tested on 50-D isotropic Gaussian blobs generated by the sklearn make_blobs function (n=300,10000,25000). A dataset of biological relevance was also selected (n=4189). ARI/F1 scores >= 0.99 were obtained for the Gaussian data when compared to the labels provided by make_blobs. ARI/F1 = 0.98 was obtained for the biological dataset when compared to sklearn-derived labels.

This affinityprop implementation was compared against the Affinity Propagation implementation contained within scikit-learn-1.0.2, and run using numpy-1.22.2 with Python 3.10.0. This analysis was completed using a Ryzen 9 3950X processor.

In all analyses, damping=0.95, convergence_iter=400, and max_iter=4000. Preference=-1000.0 for Gaussian data and -10.0 for biological data.

Time to complete analysis, scaled by the sklearn implementation, plotted against the number of cores.

Dependencies

~3MB
~58K SLoC