ngraph.forcelayout

ngraph.forcelayout

This is a force directed graph layout algorithm, that works in any dimension (2D, 3D, and above).

The library uses quad tree to speed up computation of long-distance forces.

This repository is part of ngraph family, and operates on ngraph.graph data structure.

API

All force directed algorithms are iterative. We need to perform multiple iterations of an algorithm, before graph starts looking good:

// graph is an instance of `ngraph.graph` object.
var createLayout = require('ngraph.forcelayout');
var layout = createLayout(graph);
for (var i = 0; i < ITERATIONS_COUNT; ++i) {
  layout.step();
}

// now we can ask layout where each node/link is best positioned:
graph.forEachNode(function(node) {
  console.log(layout.getNodePosition(node.id));
  // Node position is pair of x,y coordinates:
  // {x: ... , y: ... }
});

graph.forEachLink(function(link) {
  console.log(layout.getLinkPosition(link.id));
  // link position is a pair of two positions:
  // {
  //   from: {x: ..., y: ...},
  //   to: {x: ..., y: ...}
  // }
});

If you'd like to perform graph layout in space with more than two dimensions, just add one argument to this line:

let layout = createLayout(graph, {dimensions: 3}); // 3D layout
let nodePosition = layout.getNodePosition(nodeId); // has {x, y, z} attributes

Even higher dimensions are not a problem for this library:

let layout = createLayout(graph, {dimensions: 6}); // 6D layout
// Every layout with more than 3 dimensions, say N, gets additional attributes:
// c4, c5, ... cN
let nodePosition = layout.getNodePosition(nodeId); // has {x, y, z, c4, c5, c6} 

Note: Higher dimensionality comes at exponential cost of memory for every added dimension. See a performance section below for more details.

Node position and object reuse

Recently immutability became a ruling principle of javascript world. This library doesn't follow the rules, and results of getNodePosition()/getLinkPosition() will be always the same for the same node. This is true:

layout.getNodePosition(1) === layout.getNodePosition(1);

Reason for this is performance. If you are interested in storing positions somewhere else, you can do it and they still will be updated after each force directed layout iteration.

"Pin" node and initial position

Sometimes it's desirable to tell layout algorithm not to move certain nodes. This can be done with pinNode() method:

var nodeToPin = graph.getNode(nodeId);
layout.pinNode(nodeToPin, true); // now layout will not move this node

If you want to check whether node is pinned or not you can use isNodePinned() method. Here is an example how to toggle node pinning, without knowing it's original state:

var node = graph.getNode(nodeId);
layout.pinNode(node, !layout.isNodePinned(node)); // toggle it

What if you still want to move your node according to some external factor (e.g. you have initial positions, or user drags pinned node)? To do this, call setNodePosition() method:

layout.setNodePosition(nodeId, x, y);

Monitoring changes

Like many other algorithms in ngraph family, force layout monitors graph changes via graph events. It keeps layout up to date whenever graph changes:

var graph = require('ngraph.graph')(); // empty graph
var layout = require('ngraph.layout')(graph); // layout of empty graph

graph.addLink(1, 2); // create node 1 and 2, and make link between them
layout.getNodePosition(1); // returns position.

If you want to stop monitoring graph events, call dispose() method:

layout.dispose();

Physics Simulator

Simulator calculates forces acting on each body and then deduces their position via Newton's law. There are three major forces in the system:

1. Spring force keeps connected nodes together via Hooke's law
2. Each body repels each other via Coulomb's law
3. The drag force slows the entire simulation down, helping with convergence.

Body forces are calculated in n*lg(n) time with help of Barnes-Hut algorithm implemented with quadtree.

// Configure
var physicsSettings = {
  timeStep: 0.5,
  dimensions: 2,
  gravity: -12,
  theta: 0.8,
  springLength: 10,
  springCoefficient: 0.8,
  dragCoefficient: 0.9,
};

// pass it as second argument to layout:
var layout = require('ngraph.forcelayout')(graph, physicsSettings);

You can get current physics simulator from layout by checking layout.simulator property. This is a read only property.

Space occupied by graph

Finally, it's often desirable to know how much space does our graph occupy. To quickly get bounding box use getGraphRect() method:

var rect = layout.getGraphRect();
// rect.min_x, rect.min_y - left top coordinates of the bounding box
// rect.max_x, rect.max_y - right bottom coordinates of the bounding box

Manipulating bodies

This is advanced technique to get to internal state of the simulator. If you need to get a node position use regular layout.getNodePosition(nodeId) described above.

In some cases you really need to manipulate physic attributes on a body level. To get to a single body by node id:

var graph = createGraph();
graph.addLink(1, 2);

// Get body that represents node 1:
var body = layout.getBody(1);
assert(
  typeof body.pos.x === 'number' &&
  typeof body.pos.y === 'number', 'Body has position');
assert(body.mass, 'Body has mass');

To iterate over all bodies at once:

layout.forEachBody(function(body, nodeId) {
  assert(
    typeof body.pos.x === 'number' &&
    typeof body.pos.y === 'number', 'Body has position');
  assert(graph.getNode(nodeId), 'NodeId is coming from the graph');
});

Section about performance

This library is focused on performance of physical simulation. We use quad tree data structure in 2D space to approximate long distance forces, and reduce amount of required computations.

When layout is performed in higher dimensions we use analogues tree data structure. By design such tree requires to store 2^dimensions_count child nodes on each node. In practice, performing layout in 6 dimensional space on a graph with a few thousand nodes yields decent performance on modern mac book (graph can be both rendered and layed out at 60FPS rate).

Additionally, the vector algebra is optimized by a ad-hoc code generation. Essentially this means that upon first load of the library, we check the dimension of the space where you want to perform layout, and generate all required data structure to run fast in this space.

The code generation happens only once when dimension is requested. Any subsequent layouts in the same space would reuse generated codes. It is pretty fast and cool.

install

With npm do:

npm install ngraph.forcelayout

Or download from CDN:

<script src='https://unpkg.com/ngraph.forcelayout@3.0.0/dist/ngraph.forcelayout.min.js'></script>

If you download from CDN the library will be available under ngraphCreateLayout global name.

license

MIT

Feedback?

I'd totally love it! Please email me, open issue here, tweet to me, or join discussion on gitter.

If you love this library, please consider sponsoring it at https://github.com/sponsors/anvaka or at https://www.patreon.com/anvaka