void removeMatrixRow(int row, MatDoub &out) {
int dummy = -1000;
Rh.resize(Ri.nrows(),Ri.nrows());
Rh = Ri;
for(int i=0; i<Rh.nrows(); i++) {
Rh[row][i] = dummy;
Rh[i][row] = dummy;
}
datain = Rh.getMatrixArray();
data [Rh.nrows()*Rh.nrows()];
int k=0;
for(int i=0; i < Rh.nrows()*Rh.nrows(); i++) {
double ele = datain[i];
if(ele != dummy)
data[k++] = ele;
}
out = MatDoub( out.nrows(), out.nrows(), data );
}
void dump_nrmat( MatDoub &m ) {
for( int r=0; r<m.nrows(); r++ ) {
for( int c=0; c<m.ncols(); c++ ) {
printf( "%+3.2le ", m[r][c] );
}
printf( "\n" );
}
}
void removeMatrixRow( MatDoub &out ) {
int dummy = -1000;
datain = Ri.getMatrixArray();
data [Ri.nrows()*Ri.nrows()];
int k=0;
for(int i=0; i < Ri.nrows()*Ri.ncols(); i++) {
double ele = datain[i];
if(ele != dummy)
data[k++] = ele;
}
out = MatDoub( out.nrows(), out.nrows(), data );
}
/*
Utility method used by Geodesic and RandomWalk algorithms
to set-up the Modularity and Laplacian matrices.
*/
void setupMatrices() {
//--- Matrix size
int N = R.nrows();
//--- 2*m
two_m = elist.size();
//--- _norm
_norm = 1.0/(2.0*two_m);
for(int i=0; i<N; i++) {
C[i] = 0;
for(int j=0; j<N; j++) {
R[i][j] = 0.0;
A[i][j] = 0.0;
}
}
//--- Setup Matrices
//--- Store the current community each vertex belongs too
for(int i=0; i<N; i++) {
R[i][i] = n[i+1].getDegree();
C[i] = n[i+1].c;
}
//--- The Adjacency matrix, A
for(int i=0; i<elist.size(); i++) {
int ind_i = elist[i].so -1;
int ind_j = elist[i].si -1;
if( ind_i == ind_j || ind_j == ind_i )
A[ind_i][ind_j] = 1.0 * elist[i].Globalwe;
else {
A[ind_i][ind_j] = 1.0 * elist[i].Globalwe;
A[ind_j][ind_i] = 1.0 * elist[i].Globalwe;
}
}
for(int i=0; i<N; i++) {
for(int j=0; j<N; j++) {
R[i][j] = R[i][j] - A[i][j];
}
}
//--- The Modularity matrix, Bgi
for(int i=1; i<n.size(); i++) {
for(int j=1; j<n.size(); j++) {
Bgi[i-1][j-1] = A[i-1][j-1] - (n[i].getDegree() * n[j].getDegree() * _norm);
}
}
}
void addMatrixRow(MatDoub U, int row, MatDoub &out) {
int dummy = -1000;
for(int i=0; i<out.nrows(); i++) {
out[i][row] = dummy;
out[row][i] = dummy;
}
datain = U.getMatrixArray();
datainO = out.getMatrixArray();
int k = 0;
data [out.nrows()*out.nrows()];
for(int i=0; i < out.nrows()*out.ncols(); i++) {
if( datainO[i] == dummy )
data[i] = 0;
else
data[i] = datain[k++];
}
out = MatDoub( out.nrows(), out.nrows(), data );
}
/*
Utility method use by the RandomWalk algorithm to
update the Adjacency and Laplacian matrices.
*/
void upDateMatrices() {
for(int i=0; i<R.nrows(); i++) {
C[i] = 0;
for(int j=0; j<R.nrows(); j++) {
R[i][j] = 0;
}
}
//--- Setup Matrices
//--- Store the current community each vertex belongs too
for(int i=0; i<R.nrows(); i++) {
R[i][i] = n[i+1].getDegree();
C[i] = n[i+1].c;
}
//--- Update the Adjacency matrix, A
for(int i=0; i<elist.size(); i++) {
if( elist[i].removed ) {
int ind_i = elist[i].so -1;
int ind_j = elist[i].si -1;
//if edge has been removed, remove this entry
//from A.
A[ind_i][ind_j] = 0;
A[ind_j][ind_i] = 0;
}
}
//--- Update the Lapacian matrix, R
for(int i=0; i<R.nrows(); i++) {
for(int j=0; j<R.nrows(); j++) {
R[i][j] = R[i][j] - A[i][j];
}
}
}
void copyNRMatToZMat( MatDoub &m, ZMat &z ) {
// account for NR3 is rowmajor, ZMat is colmajor.
int rows = m.nrows();
int cols = m.ncols();
if( z.rows != rows || z.cols != cols ) {
z.alloc( rows, cols, zmatF64 );
}
for( int r=0; r<rows; r++ ) {
for( int c=0; c<cols; c++ ) {
z.putD( r, c, m[r][c] );
}
}
}
void removeMatrixRow( MatDoub Unr, MatDoub &outnr ) {
int dummy = -1000;
datain = Unr.getMatrixArray();
data [outnr.nrows()*outnr.nrows()];
int k=0;
for(int i=0; i < Unr.nrows()*Unr.nrows(); i++) {
double ele = datain[i];
if(ele != dummy)
data[k++] = ele;
}
outnr = MatDoub( outnr.nrows(), outnr.nrows(), data );
}
/*
Utility method used by RandomWalk algorithm to
resize the Graph Laplacian for each community, com,
within the network.
*/
void getSubMatrix(int com, vector<node> &Nodes) {
int dummy = -1000;
int rows = 0;
Rh.resize(R.nrows(), R.nrows());
Rh = R;
//--- NR style
for( int i=0; i< C.size(); i++) {
if( C[i] == com )
Nodes.push_back(node(rows++,0.0,0.0));
else {
for( int k=0; k<Rh.nrows(); k++) {
Rh[i][k] = dummy;
Rh[k][i] = dummy;
}
}
}
datain = Rh.getMatrixArray();
data [Rh.nrows()*Rh.nrows()];
int ind = 0;
for(int i=0; i < Rh.nrows()*Rh.ncols(); i++) {
double ele = datain[i];
if(ele != dummy)
data[ind++] = ele;
}
Ri.resize(rows,rows);
Ri = MatDoub( rows, rows, data );
}
/*
Calculate the split of nodes belonging to the last group of nodes
with negative eigenvector values.
*/
void splitN(MatDoub Bg, VecInt keys, int dummy, double tol) {
cout << "> In splitN method... " << endl;
int N = Bg.nrows();
MatDoub Bgii(N,N);
MatDoub Bgiii(N,N);
VecInt keysi_n (N);
//--- Starting from the group Modularity matrix Bg,
//--- resize matrices: Bgi, keysi_p, keysi_n, u and betai.
Bgiii = Bg;
int Ng = 0;
for(int i=0; i<keys.size(); i++) {
if(keys[i] != dummy) {
Ng++;
} else {
for(int k=0; k<Bgiii.nrows(); k++) {
Bgiii[i][k] = dummy;
Bgiii[k][i] = dummy;
}
}
}
keysi_n.resize(Ng);
VecInt keysi_p(Ng);
int k=0;
for(int i=0; i<keys.size(); i++) {
if(keys[i] != dummy)
keysi_n[k++] = keys[i];
}
Bgii.resize(Ng,Ng);
removeMatrixRow(Bgiii,Bgii);
Bgi.resize(Bgii.nrows(),Bgii.nrows());
//--- Calculate the Modularity matrix Bgi for the new node group
calculateB(Bgii, Bgi);
u.resize(Ng,Ng);
betai.resize(Ng);
//--- Calculate eigenvectors, and values, from Bgi...
calculateEigenVectors();
int ind = 0;
findLeadingEigenVectors(ind);
//--- Check that maximum eigenvalue is greater than the tolerance.
cout << "> max EigenValue is " << betai[ind] << " with ind " << ind << endl;
if(betai[ind] > tol ) {
//--- set up the index vectors, si and SI, for the initial split
maximiseIndexVectors(ind);
double deltaQ_old = 0.0;
double deltaQ_new = 0.0;
int cp = 0;
int cn = 0;
//--- Calculate the Spectral Modularity
deltaModularity(deltaQ_old);
cout << "> Spectral Q: " << deltaQ_old << endl;
double diff = fabs(deltaQ_old);
int count = 0;
//--- Fine tuning stage to maximum deltaModularity for the initial split
while( diff > tol ) {
modifySplit( tol, Ng );
deltaModularity(deltaQ_new);
cout << "> Modified Q: " << deltaQ_new << endl;
diff = fabs( deltaQ_new - deltaQ_old );
deltaQ_old = deltaQ_new;
}
//--- Keep recorded of maximum fine-tuned Modularity value.
specQ += deltaQ_old;
for(int i=0; i<Ng; i++) {
si[i] = si[i];
if(si[i] > 0) cp++;
else cn++;
}
if(cp < 1 || cn < 1) {
cout << "> Stop splitting. " << endl;
return;
}
int Ncomn = maxCommunity() + 1;
int Ncomp = Ncomn + 1;
cout << "> node list " << endl;
for(int i=0; i<keysi_n.size(); i++) {
if( si[i] < 0) {
keysi_n[i] = keysi_n[i];
keysi_p[i] = dummy;
n[(int)keysi_n[i]].c = Ncomn;
cout << "> Node: " << keysi_n[i] << " c = " << n[(int)keysi_n[i]].c << endl;
} else {
keysi_p[i] = keysi_n[i];
keysi_n[i] = dummy;
cout << "> Node: " << keysi_p[i] << " c = " << n[(int)keysi_p[i]].c << endl;
}
}
//--- Recursively split the group of positive eigenvector nodes
splitP(Bgii, keysi_p, dummy, tol);
//--- Recursively split the group of negative eigenvector nodes
splitN(Bgii, keysi_n, dummy, tol);
} else {
cout << "> Stop splitting. " << endl;
return ;
}
}
/*
Utility method used by the RandomWalk algorithm to
invert the Graph Laplacian and accummulate the random-path
contributions from each source-sink node pair, in
accordance with [2] (see method declarations above).
*/
void calculateRandomWalk(int c, vector<node> Nodes) {
queue<node> termNodes;
for(int i=0; i<Nodes.size(); i++)
termNodes.push(Nodes[i]);
int N = Nodes.size();
S.resize(N,1);
V.resize(N,1);
T.resize(N,N);
Rc.resize((N-1),(N-1));
Vi.resize(C.size(),1);
//--- Remove arbitrary termination ('sink') state '0'.
removeMatrixRow(0,Rc);
//--- Invert Matrix.
LUdcmp lu( Rc );
Ti.resize(Rc.nrows(),Rc.nrows());
lu.inverse( Ti );
for(int i=0; i< T.nrows(); i++) {
for(int j=0; j< T.nrows(); j++) {
T[i][j] = 0;
}
}
//--- Add back arbitrary termination ('sink') state '0'.
addMatrixRow(Ti,0,T);
while ( !termNodes.empty() ) {
node termNode = termNodes.front();
termNodes.pop();
for(int t=0; t< Nodes.size(); t++) {
//--- Take the next start ('source') state.
node startNode = Nodes[t];
if( startNode.k != termNode.k ) {
for(int i=0; i<S.nrows(); i++) {
S[i][0] = 0;
V[i][0] = 0;
Vi[i][0] = 0;
}
S[startNode.k][0] = 1;
S[termNode.k][0] = -1;
//--- V = T * S
for(int i=0; i<T.nrows(); i++) {
double sum = 0.0;
for(int j=0; j<T.nrows(); j++) {
sum += T[i][j] * S[j][0];
}
V[i][0] = sum;
}
addMatrixRows(V, c, Vi);
//--- Edge Betweenness, i.e.
//--- the currents (potential differences) alone each edge!
for(int i=0; i<elist.size(); i++) {
if( !elist[i].removed ) {
int Ni = elist[i].so-1;
int Nj = elist[i].si-1;
totallist[i+1].we += fabs(Vi[Ni][0] - Vi[Nj][0]);
}
}//elist
}//Nodes
}//startNodes
}
}
