pyrwr

Python Implementation for Random Walk with Restart (PyRWR).

Random Walk with Restart (RWR) is one of famous link analysis algorithms, which measures node-to-node proximities in arbitrary types of graphs (networks). The representative applications include various real-world graph mining tasks such as personalized node ranking, recommendation in graphs (e.g., 'who you may know'), and anormaly detection.

pyrwr aims to implement algorithms for computing RWR scores based on Power Iteration using numpy and scipy in Python. More specifically, pyrwr focuses on computing a single source RWR score vector w.r.t. a given query (seed) node, which is used for a personalized node ranking w.r.t. the querying node. Besides RWR, pyrwr supports computing Personalized PageRank (PPR) with multiple seeds and PageRank which are well-known variants of RWR.

The supported features of pyrwr are:

• Query types
  • Random Walk with Restart (RWR): personalized ranking; only a single seed is allowed
  • Personalized PageRank (PPR): personalized ranking; multiple seeds are allowed
  • PageRank: global ranking; all nodes are used as seeds
• Graph types
  • Unweighted/weighted graphs
  • Undirected/directed graphs
• GPU computation
  • If you have a gpu and set device to 'gpu', it will process your query on the gpu using pytorch. Once the input data is transformed to the gpu, it computes the query more quickly compared to a cpu. If your graph is large, but fit to the gpu memory, try this option for fast computation.

If you are interested in studying random walk based ranking models such as PageRank and RWR, please consider this hands-on tutorial (https://github.com/jinhongjung/tutorial-on-link-analysis) that provides how to correctly implement the algorithms of those models in Python and to analyze real-world networks using the ranking models.

Installation

To install this package, type the following:

cd pyrwr
pip3 install -r requirements.txt
python3 setup.py install

These will execute the installation of python modules required by this package. The name of the installed program is pyrwr. If you want to validate whether the installation is successfully finished, type the following command:

pyrwr --help

Requirements

• numpy
• scipy
• tqdm
• fire
• pandas
• torch

Usage

We provide the following simple command line usage:

pyrwr --query-type rwr --graph-type directed --input-path data/directed/sample.tsv --output-path output/scores.tsv --seeds 10982 --device cpu

This will compute an RWR score vector w.r.t. the seed node given by --seeds in the given graph specified by --input-path on a single cpu, and write the vector into the target file in --output-path. --query-type specifies the type of query, e.g., this example indicates an RWR query. The detailed format of the input and output files is described below.

Example Usage on GPU

CUDA_VISIBLE_DEVICES=1
pyrwr --query-type rwr --graph-type directed --input-path data/directed/sample.tsv --output-path output/scores.tsv --seeds 10982 --device gpu

It tries to compute the rwr query on the gpu whose device id is 1 if the gpu is available. Otherwise, an error will be raised.

Input and Output Format

Input Format

The default input of pyrwr represents the edge list of a graph with the following format (tab separated):

# format: source (int) \t target (int)
1	2
1	4
2	3
...

The above example represents an unweighted and directed graph where each line indicates an edge from source to target. In this case, the edge weight is set to 1 uniformly. The node id should be non-negative (>= 0). To vary weights edge by edge, use the following format:

# format: source (int) \t target (int) \t weight (float)
1	2	1.5	
1	4	3.5
2	3	6.0
...

Note that

• RWR is defined on positively weighted networks; thus, only positive weights are allowed.
• If there are redundant edges in an weighted network, their weights will be summed, e.g.,

# Suppose there are the following redundant edges in the original file.
1   2   3
1   2   5
---------
# Then, their weights are summed in the final adjacency matrix as follows.
1   2   8

Output Format

The default output of pyrwr contains the single source RWR score vector w.r.t. the given seed node as follows:

# format : node id \t an RWR score of the node
1   0.1232e-3
2   0.0349e-4
...

How to Use pyrwr in My Code?

The following example shows how to import pyrwr and compute an RWR query.

from pyrwr.rwr import RWR

rwr = RWR()
rwr.read_graph(input_graph, graph_type)
r = rwr.compute(seed, c, epsilon, max_iters)

Note that seed should be int. The format of the file at input_graph should follow one of the above input formats. r is a column vector (ndarray) having the RWR score vector w.r.t. seed node. The shape of r will be (n, 1) where n is the number of nodes.

For a PPR query, see the following code:

from pyrwr.ppr import PPR

ppr = PPR()
ppr.read_graph(input_graph, graph_type)
r = ppr.compute(seeds, c, epsilon, max_iters)

In this case, seeds is the list of seeds. r is the PPR score vector w.r.t. seeds. Note that the PPR vector r is used for obtaining the personalized node ranking list related to all seeds in the seeds list.

For the PageRank query, use the following snippet:

from pyrwr.pagerank import PageRank

pagerank = PageRank()
pagerank.read_graph(input_graph, graph_type)
r = pagerank.compute(c, epsilon, max_iters)

Note that for pagerank, we do not need to specify seeds since PageRank is a global ranking; thus, it automatically sets required seeds (i.e., all nodes are used as seeds).

Arguments of pyrwr

We summarize the input arguments of pyrwr in the following table:

Arguments	Query Type	Explanation	Default
query-type	common	Query type among [rwr, ppr, pagerank]	None
graph-type	common	Graph type among [directed, undirected]	None
input-path	common	Input path for a graph	None
output-path	common	Output path for storing a query result	None
seeds	rwr	A single seed node id	None
seeds	ppr	File path for seeds (str) or list of seeds (list)	[]
c	common	Restart probablity (rwr) or jumping probability (otherwise)	0.15
epsilon	common	Error tolerance for power iteration	1e-9
max-iters	common	Maximum number of iterations for power iteration	100
handles-deadend	common	If true, handles the deadend issue	True
device	common	Computing device [cpu, gpu]	cpu

The value of Query Type in the above table is one of the followings:

• common: parameter of all of rwr, ppr, and pagerank
• rwr: parameter of rwr
• ppr: parameter of ppr
• pagerank: parameter of pagerank

Note the followings:

• If you want to compute pagerank query, then do not need to specify seeds.
• For directed graphs, there might be deadend nodes whose out-degree is zero. In this case, a naive power iteration would incur leaking out scores. handles_deadend exists for such issue handling deadend nodes. With handles_deadend, you can guarantee that the sum of a score vector is 1. Otherwise, the sum would less than 1 in directed graphs. The strategy pyrwr exploits is that whenever a random surfer visits a deadend node, go back to a seed node (or one of seed nodes), and restart. See this for the detailed technique of the strategy.