rnbrw 0.1.6.3

Renewal Non-Backtracking Random Walk (RNBRW) for community detection

RNBRW

RNBRW (Renewal Non-Backtracking Random Walks) is a Python package for estimating edge-level importance in networks using random walks that restart upon cycle closure. These weights can be used to improve community detection algorithms like Louvain.

Based on:

Moradi, B., Shakeri, H., Poggi-Corradini, P., & Higgins, M.
A new method for incorporating network cyclic structures to improve community detection
arXiv:1805.07484

---

Installation

pip install rnbrw

Features

• Parallel RNBRW edge weight estimation
• Seamless integration with Louvain
• Based on Moradi-Jamei et al., 2019

Parallelization & HPC Useage

rnbrw supports parallel execution of the RNBRW simulations using joblib. This allows for efficient simulation on multi-core machines or High-Performance Computing (HPC) clusters.

You can control parallel execution using the n_jobs parameter:

• Local machine: Set n_jobs=-1 to use all available CPU cores, or specify the exact number of cores to use (e.g., n_jobs=4).

• High-Performance Computing (HPC): When running on an HPC cluster, n_jobs can be tuned according to the allocated CPUs in your job script. For best performance, align n_jobs with the number of cores requested via sbatch, qsub, or your cluster’s job scheduler.

HPC Usage

RNBRW supports both single-walk and batched multi-walk execution on HPC clusters.

• Use compute_weights(..., only_walk=True) for one walk per job (simple job arrays).
• Use walk_hole_E for batched jobs (e.g. 300 walks per job), which guarantees independence without mutating the graph.

---

Step 1: Run RNBRW Walks in Parallel Jobs

Option A – One Walk Per Job (simple arrays)

SLURM Example Here’s a basic SLURM job array script:rnbrw_job.sh

#!/bin/bash
#SBATCH --job-name=rnbrw_walk
#SBATCH --output=logs/rnbrw_%A_%a.out
#SBATCH --error=logs/rnbrw_%A_%a.err
#SBATCH --array=0-19              # 20 total jobs
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=1
#SBATCH --mem=2G
#SBATCH --time=00:10:00

module load python/3.10
source activate rnbrw-env

# Run the Python script with job array index
python run_rnbrw_batch.py $SLURM_ARRAY_TASK_ID

import sys
import numpy as np
import networkx as nx
from rnbrw.weights import compute_weights

job_id = int(sys.argv[1])
seed = 1000 + job_id

# Load the graph (shared across all jobs)
G = nx.read_gpickle("mygraph.gpickle")

# Single walk using only_walk mode
G = compute_weights(G, nsim=1, seed=seed, only_walk=True)

# Extract edge counts
m = G.number_of_edges()
T = np.zeros(m)
for u, v in G.edges():
    T[G[u][v]['enum']] = G[u][v]['ret']

np.save(f"T_partial_{job_id}.npy", T)
 

Option A – Batched Walks Per Job (e.g. 300 walks/job)

import sys, pickle import numpy as np from rnbrw.weights.rnbrw import walk_hole_E # low-level primitive

job_id = int(sys.argv[1]) walks_per_job = 300 seeds = [1000 + job_id * walks_per_job + i for i in range(walks_per_job)]

Load the graph once

with open("mygraph.gpickle", "rb") as f: G = pickle.load(f)

m = G.number_of_edges() T = np.zeros(m)

for s in seeds: T += walk_hole_E(G, seed=s) # independent walks, no mutation

np.save(f"T_partial_{job_id}.npy", T)

Step 2 – Aggregate outputs (on head node):

import numpy as np

T_total = sum(np.load(f"T_partial_{i}.npy") for i in range(num_jobs))

Step 3 – Assign Weights to Graph

import networkx as nx
import numpy as np
from rnbrw.utils import assign_rnbrw_weights

G = nx.read_gpickle("mygraph.gpickle")
T_total = np.load("T_total.npy")

# Assign raw + normalized weights to the graph
G = assign_rnbrw_weights(G, T_total)

Step 4: Run Louvain

from rnbrw.community import detect_communities_louvain

partition = detect_communities_louvain(G, weight_attr='ret_n')

This makes rnbrw especially suitable for research environments where cycles and edge roles must be computed across very large networks.

Local Usage

Use compute_weights directly with multi-threading:

import networkx as nx
from rnbrw.weights import compute_weights
from rnbrw.community import detect_communities_louvain

# Create or load a graph
G = nx.karate_club_graph()

# Compute RNBRW weights
# Recommendation: nsim should be at least the number of edges in G
G = compute_weights(G, nsim=G.number_of_edges(), n_jobs=4)

# Edge weights (normalized)
weights = [G[u][v]['ret_n'] for u, v in G.edges()]

# Detect communities
from rnbrw.community import detect_communities_louvain

partition = detect_communities_louvain(G, weight_attr='ret_n')

API Reference

G_weighted = compute_weights(
    G,               # networkx.Graph
    nsim=None,       # Optional[int], defaults to factor * m (num edges)
    factor=1.0,      # float, multiplies number of edges to compute default nsim
    seed=None,       # Optional[int], random seed for reproducibility
    n_jobs=1,        # int, number of parallel jobs (-1 = all CPUs)
    init_weight=0.001,# float, initial placeholder edge weights
    only_walk=False  # bool, run single walk (no aggregation) for HPC
)

Simulates RNBRW on graph G to assign edge importance scores as weights.

Parameters for compute_weights

Parameter	Type	Default	Description
G	networkx.Graph	required	Input undirected graph.
nsim	int or None	None	Number of RNBRW simulations. If None, it defaults to factor × m, where m is the number of edges.
factor	float	1.0	Scaling factor to set nsim dynamically based on graph size (m).
n_jobs	int	1	Number of parallel jobs (-1 uses all available CPUs).
seed	int or None	None	Random seed for reproducibility. Each simulation is seeded with seed + i.
init_weight	float	0.01	Initial placeholder weight for each edge before running RNBRW.
only_walk	bool	False	If True, performs a single walk without aggregating weights (for HPC use).

Notes

• Recommended: For stable and convergent RNBRW edge weights, set nsim approximately equal to m, the number of edges in the graph (e.g., nsim ≈ m).
• If only_walk=True, the function will return the walk output without updating graph weights — useful for splitting across HPC batch jobs manually.

detect_communities_louvain(G, weight_attr='ret_n')

Runs Louvain on G using edge weights.

Parameter	Type	Description
G	networkx.Graph	Weighted input graph
weight_attr	str	Edge weight attribute used for Louvain (default = 'ret_n')

|

normalize_edge_weights(G, weight='ret')

Normalizes the weights to sum to 1 across all edges.

Parameter	Type	Description
G	networkx.Graph	Graph whose edge weights are to be normalized
weight	str	Edge attribute to normalize (default = 'ret')

If you are running RNBRW simulations in parallel (e.g., on an HPC cluster), you can aggregate the walk counts and assign RNBRW weights manually using the function below.

# After accumulating total walk counts across jobs:
G_weighted = assign_rnbrw_weights(G, T_total)
# G must have 'enum' attributes on edges

Parameter	Type	Description
G	networkx.Graph	Input graph with edges having 'enum' attribute for indexing
T_total	np.ndarray	Array of edge hit counts from RNBRW simulations (same order as 'enum')

Use this when you aggregate RNBRW cycle counts manually and want to assign weights post-hoc.

Citation

If you use this package in your research, please cite:

@article{moradi2018new, title={A new method for incorporating network cyclic structures to improve community detection}, author={Moradi, Behnaz and Shakeri, Heman and Poggi-Corradini, Pietro and Higgins, Michael}, journal={arXiv preprint arXiv:1805.07484}, year={2018} } Or use the “Cite this repository” button above.

License

This project is licensed under the MIT License © 2025 Behnaz Moradi-Jamei.

Documentation

Full documentation is available at Read the Docs.