cytoscape/src/collection/algorithms/markov-clustering.js

// Implemented by Zoe Xi @zoexi for GSOC 2016
// https://github.com/cytoscape/cytoscape.js-markov-cluster

// Implemented from Stijn van Dongen's (author of MCL algorithm) documentation: http://micans.org/mcl/
// and lecture notes: https://www.cs.ucsb.edu/~xyan/classes/CS595D-2009winter/MCL_Presentation2.pdf

import * as util from '../../util';

/* eslint-disable no-unused-vars */
let defaults = util.defaults({
  expandFactor: 2,      // affects time of computation and cluster granularity to some extent: M * M
  inflateFactor: 2,     // affects cluster granularity (the greater the value, the more clusters): M(i,j) / E(j)
  multFactor: 1,        // optional self loops for each node. Use a neutral value to improve cluster computations.
  maxIterations: 20,    // maximum number of iterations of the MCL algorithm in a single run
  attributes: [         // attributes/features used to group nodes, ie. similarity values between nodes
    function(edge) {
      return 1;
    }
  ]
});
/* eslint-enable */

let setOptions = ( options ) => defaults( options );

/* eslint-disable no-unused-vars, no-console */
if( process.env.NODE_ENV !== 'production' ){
  var printMatrix = function( M ) { // used for debugging purposes only
    let n = Math.sqrt(M.length);
    for ( let i = 0; i < n; i++ ) {
      let row = '';
      for ( let j = 0; j < n; j++ ) {
        row += Number(M[i*n+j]).toFixed(3) + ' ';
      }
      console.log(row);
    }
    console.log('');
  };
}
/* eslint-enable */

let getSimilarity = function( edge, attributes ) {
  let total = 0;
  for ( let i = 0; i < attributes.length; i++ ) {
    total += attributes[i]( edge );
  }
  return total;
};

let addLoops = function( M, n, val ) {
  for (let i = 0; i < n; i++) {
    M[i * n + i] = val;
  }
};

let normalize = function( M, n ) {
  let sum;
  for ( let col = 0; col < n; col++ ) {
    sum = 0;
    for ( let row = 0; row < n; row++ ) {
      sum += M[row * n + col];
    }
    for ( let row = 0; row < n; row++ ) {
      M[row * n + col] = M[row * n + col] / sum;
    }
  }
};

// TODO: blocked matrix multiplication?
let mmult = function( A, B, n ) {
  let C = new Array( n * n );

  for ( let i = 0; i < n; i++ ) {
    for ( let j = 0; j < n; j++ ) {
      C[i * n + j] = 0;
    }

    for ( let k = 0; k < n; k++ ) {
      for ( let j = 0; j < n; j++ ) {
        C[i * n + j] += A[i * n + k] * B[k * n + j];
      }
    }
  }
  return C;
};

let expand = function( M, n, expandFactor /** power **/ ) {
  let _M = M.slice(0);

  for ( let p = 1; p < expandFactor; p++ ) {
    M = mmult( M, _M, n );
  }
  return M;
};

let inflate = function( M, n, inflateFactor /** r **/ ) {
  let _M = new Array( n * n );

  // M(i,j) ^ inflatePower
  for ( let i = 0; i < n * n; i++ ) {
    _M[i] = Math.pow( M[i], inflateFactor );
  }

  normalize( _M, n );

  return _M;
};

let hasConverged = function( M, _M, n2, roundFactor ) {
  // Check that both matrices have the same elements (i,j)
  for ( let i = 0; i < n2; i++ ) {
    let v1 = Math.round( M[i] * Math.pow(10, roundFactor) ) / Math.pow(10, roundFactor); // truncate to 'roundFactor' decimal places
    let v2 = Math.round( _M[i] * Math.pow(10, roundFactor) ) / Math.pow(10, roundFactor);

    if ( v1 !== v2 ) {
      return false;
    }
  }
  return true;
};

let assign = function( M, n, nodes, cy ) {
  let clusters = [];

  for ( let i = 0; i < n; i++ ) {
    let cluster = [];
    for ( let j = 0; j < n; j++ ) {
      // Row-wise attractors and elements that they attract belong in same cluster
      if ( Math.round( M[i * n + j] * 1000 ) / 1000 > 0 ) {
        cluster.push( nodes[j] );
      }
    }
    if ( cluster.length !== 0 ) {
      clusters.push( cy.collection(cluster) );
    }
  }
  return clusters;
};

let isDuplicate = function( c1, c2 ) {
  for (let i = 0; i < c1.length; i++) {
    if (!c2[i] || c1[i].id() !== c2[i].id()) {
      return false;
    }
  }
  return true;
};

let removeDuplicates = function( clusters ) {

  for (let i = 0; i < clusters.length; i++) {
    for (let j = 0; j < clusters.length; j++) {
      if (i != j && isDuplicate(clusters[i], clusters[j])) {
        clusters.splice(j, 1);
      }
    }
  }
  return clusters;
};

let markovClustering = function( options ) {
  let nodes = this.nodes();
  let edges = this.edges();
  let cy = this.cy();
  
  // Set parameters of algorithm:
  let opts = setOptions( options );

  // Map each node to its position in node array
  let id2position = {};
  for( let i = 0; i < nodes.length; i++ ){
    id2position[ nodes[i].id() ] = i;
  }

  // Generate stochastic matrix M from input graph G (should be symmetric/undirected)
  let n = nodes.length, n2 = n * n;
  let M = new Array( n2 ), _M;
  for ( let i = 0; i < n2; i++ ) {
    M[i] = 0;
  }

  for ( let e = 0; e < edges.length; e++ ) {
    let edge = edges[e];
    let i = id2position[ edge.source().id() ];
    let j = id2position[ edge.target().id() ];

    let sim = getSimilarity( edge, opts.attributes );

    M[i * n + j] += sim; // G should be symmetric and undirected
    M[j * n + i] += sim;
  }

  // Begin Markov cluster algorithm

  // Step 1: Add self loops to each node, ie. add multFactor to matrix diagonal
  addLoops( M, n, opts.multFactor );

  // Step 2: M = normalize( M );
  normalize( M, n );

  let isStillMoving = true;
  let iterations = 0;
  while ( isStillMoving && iterations < opts.maxIterations ) {

    isStillMoving = false;

    // Step 3:
    _M = expand( M, n, opts.expandFactor );

    // Step 4:
    M = inflate( _M, n, opts.inflateFactor );

    // Step 5: check to see if ~steady state has been reached
    if ( ! hasConverged( M, _M, n2, 4 ) ) {
      isStillMoving = true;
    }

    iterations++;
  }

  // Build clusters from matrix
  let clusters = assign( M, n, nodes, cy );

  // Remove duplicate clusters due to symmetry of graph and M matrix
  clusters = removeDuplicates( clusters );

  return clusters;
};

export default {
  markovClustering,
  mcl: markovClustering
};