miazzy / ngraph.pagerank

PageRank calculation for ngraph.graph

ngraph.pagerank

PageRank algorithm implementation in JavaScript. This module is part of ngraph family.

usage

Let's compute PageRank for a simple graph with two nodes, one edge:

var graph = require('ngraph.graph')();
graph.addLink(1, 2);

var pagerank = require('ngraph.pagerank');
var rank = pagerank(graph);

This code will compute PageRank for two nodes:

{
  "1": 0.350877,
  "2": 0.649123
}

configuring

The PageRank algorithm allows you to specify a probability at any step that a person will continue clicking outgoing links. This probability in some literature is called a dumping factor, and is recommended to be set between 0.80 and 0.90.

To configure this probability use the second, optional argument of the pagerank() function:

// by default this value is 0.85. Bump it to 0.9:
var internalJumpProbability = 0.90;
var rank = pagerank(graph, internalJumpProbability);

Current implementation uses approximate solution for eigenvector problem. To specify precision level use the last optional argument:

var internalJumpProbability = 0.85;
// by default it's set to 0.005, let's increase it:
var precision = 0.00001;
var rank = pagerank(graph, internalJumpProbability, precision);

precision will affect algorithm performance and serves as an exit criteria:

|r(t) - r(t - 1)| < precision

Here r(t) is eigenvector (or pagerank of a graph) at time step t.

performance

The focus of this module is to be very fast. I tried multiple approaches, including

• Easy to read code with plain old javascript objects
• Approach with typed arrays, where objects are stored into flat array
• C++ version of the code, compiled into asm.js and extracted into separate module

So far approach with typed array gives the fastest results in v8/node.js 0.12: 43 ops/sec. asm.js version is the fastest when executed inside spider monkey (firefox) with 50 ops/sec. Unfortunately asm.js version gives terrible results in iojs 1.5 (around 20 ops/sec), and while performs at 47 ops/sec in node.js 0.12 the deviation is too big (around 7%) to call it stable. I'm frankly a little bit lost and not sure why asm.js gives such poor results in v8. So currently sticking with approach with typed arrays.

demo

A small demo is available here. It computes PageRank for a graph from Wikipedia and then renders it with force based layout.

install

With npm do:

npm install ngraph.pagerank

license

MIT

TODO:

Implement topic-specific rank?