Multilevel Multilevel_new(SparseMatrix A0, SparseMatrix D0, real *node_weights, Multilevel_control ctrl){
/* A: the weighting matrix. D: the distance matrix, could be NULL. If not null, the two matrices must have the same sparsity pattern */
Multilevel grid;
SparseMatrix A = A0, D = D0;
if (!SparseMatrix_is_symmetric(A, FALSE) || A->type != MATRIX_TYPE_REAL){
A = SparseMatrix_get_real_adjacency_matrix_symmetrized(A);
}
if (D && (!SparseMatrix_is_symmetric(D, FALSE) || D->type != MATRIX_TYPE_REAL)){
D = SparseMatrix_symmetrize_nodiag(D, FALSE);
}
grid = Multilevel_init(A, D, node_weights);
grid = Multilevel_establish(grid, ctrl);
if (A != A0) grid->delete_top_level_A = TRUE;/* be sure to clean up later */
return grid;
}
void uniform_stress(int dim, SparseMatrix A, real *x, int *flag){
UniformStressSmoother sm;
real lambda0 = 10.1, M = 100, scaling = 1.;
real res;
int maxit = 300, samepoint = TRUE, i, k, n = A->m;
SparseMatrix B = NULL;
*flag = 0;
/* just set random initial for now */
for (i = 0; i < dim*n; i++) {
x[i] = M*drand();
}
/* make sure x is not all at the same point */
for (i = 1; i < n; i++){
for (k = 0; k < dim; k++) {
if (ABS(x[0*dim+k] - x[i*dim+k]) > MACHINEACC){
samepoint = FALSE;
i = n;
break;
}
}
}
if (samepoint){
srand(1);
#ifdef DEBUG_PRINT
fprintf(stderr,"input coordinates to uniform_stress are the same, use random coordinates as initial input");
#endif
for (i = 0; i < dim*n; i++) x[i] = M*drand();
}
B = get_distance_matrix(A, scaling);
assert(SparseMatrix_is_symmetric(B, FALSE));
sm = UniformStressSmoother_new(dim, B, x, 1000000*lambda0, M, flag);
res = UniformStressSmoother_smooth(sm, dim, x, maxit);
UniformStressSmoother_delete(sm);
sm = UniformStressSmoother_new(dim, B, x, 10000*lambda0, M, flag);
res = UniformStressSmoother_smooth(sm, dim, x, maxit);
UniformStressSmoother_delete(sm);
sm = UniformStressSmoother_new(dim, B, x, 100*lambda0, M, flag);
res = UniformStressSmoother_smooth(sm, dim, x, maxit);
UniformStressSmoother_delete(sm);
sm = UniformStressSmoother_new(dim, B, x, lambda0, M, flag);
res = UniformStressSmoother_smooth(sm, dim, x, maxit);
UniformStressSmoother_delete(sm);
scale_to_box(0,0,7*70,10*70,A->m,dim,x);;
SparseMatrix_delete(B);
}
Multilevel_MQ_Clustering Multilevel_MQ_Clustering_init(SparseMatrix A, int level) {
Multilevel_MQ_Clustering grid;
int n = A->n, i;
int *matching;
assert(A->type == MATRIX_TYPE_REAL);
assert(SparseMatrix_is_symmetric(A, FALSE));
if (!A) return NULL;
assert(A->m == n);
grid = MALLOC(sizeof(struct Multilevel_MQ_Clustering_struct));
grid->level = level;
grid->n = n;
grid->A = A;
grid->P = NULL;
grid->R = NULL;
grid->next = NULL;
grid->prev = NULL;
grid->delete_top_level_A = FALSE;
matching = grid->matching = MALLOC(sizeof(real)*(n));
grid->deg_intra = NULL;
grid->dout = NULL;
grid->wgt = NULL;
if (level == 0) {
real mq = 0, mq_in, mq_out;
int n = A->n, ncluster;
real *deg_intra, *wgt, *dout;
grid->deg_intra = MALLOC(sizeof(real)*(n));
deg_intra = grid->deg_intra;
grid->wgt = MALLOC(sizeof(real)*n);
wgt = grid->wgt;
for (i = 0; i < n; i++) {
deg_intra[i] = 0;
wgt[i] = 1.;
}
for (i = 0; i < n; i++) matching[i] = i;
mq = get_mq(A, matching, &ncluster, &mq_in, &mq_out, &dout);
fprintf(stderr,"ncluster = %d, mq = %f\n", ncluster, mq);
grid->mq = mq;
grid->mq_in = mq_in;
grid->mq_out = mq_out;
grid->dout = dout;
grid->ncluster = ncluster;
}
return grid;
}
Multilevel Multilevel_new(SparseMatrix A0, real *node_weights, Multilevel_control ctrl){
Multilevel grid;
SparseMatrix A = A0;
if (!SparseMatrix_is_symmetric(A, FALSE) || A->type != MATRIX_TYPE_REAL){
A = SparseMatrix_get_real_adjacency_matrix_symmetrized(A);
}
grid = Multilevel_init(A, node_weights);
grid = Multilevel_establish(grid, ctrl);
if (A != A0) grid->delete_top_level_A = TRUE;/* be sure to clean up later */
return grid;
}
void stress_model_core(int dim, SparseMatrix B, real **x, int edge_len_weighted, int maxit_sm, real tol, int *flag){
int m;
SparseStressMajorizationSmoother sm;
real lambda = 0;
/*int maxit_sm = 1000, i; tol = 0.001*/
int i;
SparseMatrix A = B;
if (!SparseMatrix_is_symmetric(A, FALSE) || A->type != MATRIX_TYPE_REAL){
if (A->type == MATRIX_TYPE_REAL){
A = SparseMatrix_symmetrize(A, FALSE);
A = SparseMatrix_remove_diagonal(A);
} else {
A = SparseMatrix_get_real_adjacency_matrix_symmetrized(A);
}
}
A = SparseMatrix_remove_diagonal(A);
*flag = 0;
m = A->m;
if (!x) {
*x = MALLOC(sizeof(real)*m*dim);
srand(123);
for (i = 0; i < dim*m; i++) (*x)[i] = drand();
}
if (edge_len_weighted){
sm = SparseStressMajorizationSmoother_new(A, dim, lambda, *x, WEIGHTING_SCHEME_SQR_DIST, TRUE);/* do not under weight the long distances */
//sm = SparseStressMajorizationSmoother_new(A, dim, lambda, *x, WEIGHTING_SCHEME_INV_DIST, TRUE);/* do not under weight the long distances */
} else {
sm = SparseStressMajorizationSmoother_new(A, dim, lambda, *x, WEIGHTING_SCHEME_NONE, TRUE);/* weight the long distances */
}
if (!sm) {
*flag = -1;
goto RETURN;
}
sm->tol_cg = 0.1; /* we found that there is no need to solve the Laplacian accurately */
sm->scheme = SM_SCHEME_STRESS;
SparseStressMajorizationSmoother_smooth(sm, dim, *x, maxit_sm, tol);
for (i = 0; i < dim*m; i++) {
(*x)[i] /= sm->scaling;
}
SparseStressMajorizationSmoother_delete(sm);
RETURN:
if (A != B) SparseMatrix_delete(A);
}
static void ideal_distance_avoid_overlap(int dim, SparseMatrix A, real *x, real *width, real *ideal_distance, real *tmax, real *tmin){
/* if (x1>x2 && y1 > y2) we want either x1 + t (x1-x2) - x2 > (width1+width2), or y1 + t (y1-y2) - y2 > (height1+height2),
hence t = MAX(expandmin, MIN(expandmax, (width1+width2)/(x1-x2) - 1, (height1+height2)/(y1-y2) - 1)), and
new ideal distance = (1+t) old_distance. t can be negative sometimes.
The result ideal distance is set to negative if the edge needs shrinking
*/
int i, j, jj;
int *ia = A->ia, *ja = A->ja;
real dist, dx, dy, wx, wy, t;
real expandmax = 1.5, expandmin = 1;
*tmax = 0;
*tmin = 1.e10;
assert(SparseMatrix_is_symmetric(A, FALSE));
for (i = 0; i < A->m; i++){
for (j = ia[i]; j < ia[i+1]; j++){
jj = ja[j];
if (jj == i) continue;
dist = distance(x, dim, i, jj);
dx = ABS(x[i*dim] - x[jj*dim]);
dy = ABS(x[i*dim+1] - x[jj*dim+1]);
wx = width[i*dim]+width[jj*dim];
wy = width[i*dim+1]+width[jj*dim+1];
if (dx < MACHINEACC*wx && dy < MACHINEACC*wy){
ideal_distance[j] = sqrt(wx*wx+wy*wy);
*tmax = 2;
} else {
if (dx < MACHINEACC*wx){
t = wy/dy;
} else if (dy < MACHINEACC*wy){
t = wx/dx;
} else {
t = MIN(wx/dx, wy/dy);
}
if (t > 1) t = MAX(t, 1.001);/* no point in things like t = 1.00000001 as this slow down convergence */
*tmax = MAX(*tmax, t);
*tmin = MIN(*tmin, t);
t = MIN(expandmax, t);
t = MAX(expandmin, t);
if (t > 1) {
ideal_distance[j] = t*dist;
} else {
ideal_distance[j] = -t*dist;
}
}
}
}
return;
}
static void get_neighborhood_precision_recall(char *outfile, SparseMatrix A0, real *ideal_dist_matrix, real *dist_matrix){
SparseMatrix A = A0;
int i, j, k, n = A->m;
// int *ia, *ja;
int *g_order = NULL, *p_order = NULL;/* ordering using graph/physical distance */
real *gdist, *pdist, radius;
int np_neighbors;
int ng_neighbors; /*number of (graph theoretical) neighbors */
real node_dist;/* distance of a node to the center node */
real true_positive;
real recall;
FILE *fp;
fp = fopen(outfile,"w");
if (!SparseMatrix_is_symmetric(A, FALSE)){
A = SparseMatrix_symmetrize(A, FALSE);
}
// ia = A->ia;
// ja = A->ja;
for (k = 5; k <= 50; k+= 5){
recall = 0;
for (i = 0; i < n; i++){
gdist = &(ideal_dist_matrix[i*n]);
vector_ordering(n, gdist, &g_order, TRUE);
pdist = &(dist_matrix[i*n]);
vector_ordering(n, pdist, &p_order, TRUE);
ng_neighbors = MIN(n-1, k); /* set the number of closest neighbor in the graph space to consider, excluding the node itself */
np_neighbors = ng_neighbors;/* set the number of closest neighbor in the embedding to consider, excluding the node itself */
radius = pdist[p_order[np_neighbors]];
true_positive = 0;
for (j = 1; j <= ng_neighbors; j++){
node_dist = pdist[g_order[j]];/* the phisical distance for j-th closest node (in graph space) */
if (node_dist <= radius) true_positive++;
}
recall += true_positive/np_neighbors;
}
recall /= n;
fprintf(fp,"%d %f\n", k, recall);
}
fprintf(stderr,"wrote precision/recall in file %s\n", outfile);
fclose(fp);
if (A != A0) SparseMatrix_delete(A);
FREE(g_order); FREE(p_order);
}
Multilevel_MQ_Clustering Multilevel_MQ_Clustering_new(SparseMatrix A0, int maxcluster) {
/* maxcluster is used to specify the maximum number of cluster desired, e.g., maxcluster=10 means that a maximum of 10 clusters
is desired. this may not always be realized, and mq may be low when this is specified. Default: maxcluster = 0 */
Multilevel_MQ_Clustering grid;
SparseMatrix A = A0;
if (maxcluster <= 0) maxcluster = A->m;
if (!SparseMatrix_is_symmetric(A, FALSE) || A->type != MATRIX_TYPE_REAL) {
A = SparseMatrix_get_real_adjacency_matrix_symmetrized(A);
}
grid = Multilevel_MQ_Clustering_init(A, 0);
grid = Multilevel_MQ_Clustering_establish(grid, maxcluster);
if (A != A0) grid->delete_top_level_A = TRUE;/* be sure to clean up later */
return grid;
}
void stress_model(int dim, SparseMatrix B, real **x, int maxit_sm, real tol, int *flag){
int m;
SparseStressMajorizationSmoother sm;
real lambda = 0;
/*int maxit_sm = 1000, i; tol = 0.001*/
int i;
SparseMatrix A = B;
if (!SparseMatrix_is_symmetric(A, FALSE) || A->type != MATRIX_TYPE_REAL){
if (A->type == MATRIX_TYPE_REAL){
A = SparseMatrix_symmetrize(A, FALSE);
A = SparseMatrix_remove_diagonal(A);
} else {
A = SparseMatrix_get_real_adjacency_matrix_symmetrized(A);
}
}
A = SparseMatrix_remove_diagonal(A);
*flag = 0;
m = A->m;
if (!x) {
*x = MALLOC(sizeof(real)*m*dim);
srand(123);
for (i = 0; i < dim*m; i++) (*x)[i] = drand();
}
sm = SparseStressMajorizationSmoother_new(A, dim, lambda, *x, WEIGHTING_SCHEME_NONE);/* do not under weight the long distances */
if (!sm) {
*flag = -1;
goto RETURN;
}
SparseStressMajorizationSmoother_smooth(sm, dim, *x, maxit_sm, 0.001);
for (i = 0; i < dim*m; i++) {
(*x)[i] /= sm->scaling;
}
SparseStressMajorizationSmoother_delete(sm);
RETURN:
if (A != B) SparseMatrix_delete(A);
}
/* ================================ spring and spring-electrical based smoother ================ */
SpringSmoother SpringSmoother_new(SparseMatrix A, int dim, spring_electrical_control ctrl, real *x){
SpringSmoother sm;
int i, j, k, l, m = A->m, *ia = A->ia, *ja = A->ja, *id, *jd;
int *mask, nz;
real *d, *dd;
real *avg_dist;
SparseMatrix ID = NULL;
assert(SparseMatrix_is_symmetric(A, FALSE));
ID = ideal_distance_matrix(A, dim, x);
dd = (real*) ID->a;
sm = N_GNEW(1,struct SpringSmoother_struct);
mask = N_GNEW(m,int);
avg_dist = N_GNEW(m,real);
for (i = 0; i < m ;i++){
avg_dist[i] = 0;
nz = 0;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
avg_dist[i] += distance(x, dim, i, ja[j]);
nz++;
}
assert(nz > 0);
avg_dist[i] /= nz;
}
for (i = 0; i < m; i++) mask[i] = -1;
nz = 0;
for (i = 0; i < m; i++){
mask[i] = i;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (mask[k] != i){
mask[k] = i;
nz++;
}
}
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
for (l = ia[k]; l < ia[k+1]; l++){
if (mask[ja[l]] != i){
mask[ja[l]] = i;
nz++;
}
}
}
}
sm->D = SparseMatrix_new(m, m, nz, MATRIX_TYPE_REAL, FORMAT_CSR);
if (!(sm->D)){
SpringSmoother_delete(sm);
return NULL;
}
id = sm->D->ia; jd = sm->D->ja;
d = (real*) sm->D->a;
id[0] = 0;
nz = 0;
for (i = 0; i < m; i++){
mask[i] = i+m;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (mask[k] != i+m){
mask[k] = i+m;
jd[nz] = k;
d[nz] = (avg_dist[i] + avg_dist[k])*0.5;
d[nz] = dd[j];
nz++;
}
}
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
for (l = ia[k]; l < ia[k+1]; l++){
if (mask[ja[l]] != i+m){
mask[ja[l]] = i+m;
jd[nz] = ja[l];
d[nz] = (avg_dist[i] + 2*avg_dist[k] + avg_dist[ja[l]])*0.5;
d[nz] = dd[j]+dd[l];
nz++;
}
}
}
id[i+1] = nz;
}
sm->D->nz = nz;
sm->ctrl = spring_electrical_control_new();
*(sm->ctrl) = *ctrl;
sm->ctrl->random_start = FALSE;
sm->ctrl->multilevels = 1;
sm->ctrl->step /= 2;
sm->ctrl->maxiter = 20;
FREE(mask);
FREE(avg_dist);
SparseMatrix_delete(ID);
return sm;
}
UniformStressSmoother UniformStressSmoother_new(int dim, SparseMatrix A, real *x, real alpha, real M, int *flag){
UniformStressSmoother sm;
int i, j, k, m = A->m, *ia = A->ia, *ja = A->ja, *iw, *jw, *id, *jd;
int nz;
real *d, *w, *a = (real*) A->a;
real diag_d, diag_w, dist, epsilon = 0.01;
assert(SparseMatrix_is_symmetric(A, FALSE));
sm = MALLOC(sizeof(struct StressMajorizationSmoother_struct));
sm->data = NULL;
sm->scheme = SM_SCHEME_UNIFORM_STRESS;
sm->lambda = NULL;
sm->data = MALLOC(sizeof(real)*2);
((real*) sm->data)[0] = alpha;
((real*) sm->data)[1] = M;
sm->data_deallocator = FREE;
/* Lw and Lwd have diagonals */
sm->Lw = SparseMatrix_new(m, m, A->nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
sm->Lwd = SparseMatrix_new(m, m, A->nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
iw = sm->Lw->ia; jw = sm->Lw->ja;
id = sm->Lwd->ia; jd = sm->Lwd->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
if (!(sm->Lw) || !(sm->Lwd)) {
StressMajorizationSmoother_delete(sm);
return NULL;
}
iw = sm->Lw->ia; jw = sm->Lw->ja;
id = sm->Lwd->ia; jd = sm->Lwd->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
iw[0] = id[0] = 0;
nz = 0;
for (i = 0; i < m; i++){
diag_d = diag_w = 0;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (k != i){
dist = MAX(ABS(a[j]), epsilon);
jd[nz] = jw[nz] = k;
w[nz] = -1/(dist*dist);
w[nz] = -1.;
d[nz] = w[nz]*dist;
diag_w += w[nz];
diag_d += d[nz];
nz++;
}
}
jd[nz] = jw[nz] = i;
w[nz] = -diag_w;
d[nz] = -diag_d;
nz++;
iw[i+1] = nz;
id[i+1] = nz;
}
sm->Lw->nz = nz;
sm->Lwd->nz = nz;
return sm;
}
OverlapSmoother OverlapSmoother_new(SparseMatrix A, int m,
int dim, real lambda0, real *x, real *width, int include_original_graph, int neighborhood_only,
real *max_overlap, real *min_overlap,
int edge_labeling_scheme, int n_constr_nodes, int *constr_nodes, SparseMatrix A_constr, int shrink
){
OverlapSmoother sm;
int i, j, k, *iw, *jw, jdiag;
SparseMatrix B;
real *lambda, *d, *w, diag_d, diag_w, dist;
assert((!A) || SparseMatrix_is_symmetric(A, FALSE));
sm = GNEW(struct OverlapSmoother_struct);
sm->scheme = SM_SCHEME_NORMAL;
if (constr_nodes && n_constr_nodes > 0 && edge_labeling_scheme != ELSCHEME_NONE){
sm->scheme = SM_SCHEME_NORMAL_ELABEL;
sm->data = relative_position_constraints_new(A_constr, edge_labeling_scheme, n_constr_nodes, constr_nodes);
sm->data_deallocator = relative_position_constraints_delete;
} else {
sm->data = NULL;
}
sm->tol_cg = 0.01;
sm->maxit_cg = sqrt((double) A->m);
lambda = sm->lambda = N_GNEW(m,real);
for (i = 0; i < m; i++) sm->lambda[i] = lambda0;
B= call_tri(m, dim, x);
if (!neighborhood_only){
SparseMatrix C, D;
C = get_overlap_graph(dim, m, x, width, 0);
D = SparseMatrix_add(B, C);
SparseMatrix_delete(B);
SparseMatrix_delete(C);
B = D;
}
if (include_original_graph){
sm->Lw = SparseMatrix_add(A, B);
SparseMatrix_delete(B);
} else {
sm->Lw = B;
}
sm->Lwd = SparseMatrix_copy(sm->Lw);
#ifdef DEBUG
{
FILE *fp;
fp = fopen("/tmp/111","w");
export_embedding(fp, dim, sm->Lwd, x, NULL);
fclose(fp);
}
#endif
if (!(sm->Lw) || !(sm->Lwd)) {
OverlapSmoother_delete(sm);
return NULL;
}
assert((sm->Lwd)->type == MATRIX_TYPE_REAL);
ideal_distance_avoid_overlap(dim, sm->Lwd, x, width, (real*) (sm->Lwd->a), max_overlap, min_overlap);
/* no overlap at all! */
if (*max_overlap < 1 && shrink){
real scale_sta = MIN(1, *max_overlap*1.0001), scale_sto = 1;
if (Verbose) fprintf(stderr," no overlap (overlap = %f), rescale to shrink\n", *max_overlap - 1);
scale_sta = overlap_scaling(dim, m, x, width, scale_sta, scale_sto, 0.0001, 15);
*max_overlap = 1;
goto RETURN;
}
iw = sm->Lw->ia; jw = sm->Lw->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
for (i = 0; i < m; i++){
diag_d = diag_w = 0;
jdiag = -1;
for (j = iw[i]; j < iw[i+1]; j++){
k = jw[j];
if (k == i){
jdiag = j;
continue;
}
if (d[j] > 0){/* those edges that needs expansion */
w[j] = -100/d[j]/d[j];
/*w[j] = 100/d[j]/d[j];*/
} else {/* those that needs shrinking is set to negative in ideal_distance_avoid_overlap */
/*w[j] = 1/d[j]/d[j];*/
w[j] = -1/d[j]/d[j];
d[j] = -d[j];
}
dist = d[j];
diag_w += w[j];
d[j] = w[j]*dist;
diag_d += d[j];
}
lambda[i] *= (-diag_w);/* alternatively don't do that then we have a constant penalty term scaled by lambda0 */
assert(jdiag >= 0);
w[jdiag] = -diag_w + lambda[i];
d[jdiag] = -diag_d;
}
RETURN:
return sm;
}
static real get_mq(SparseMatrix A, int *assignment, int *ncluster0, real *mq_in0, real *mq_out0, real **dout0) {
/* given a symmetric matrix representation of a graph and an assignment of nodes into clusters, calculate the modularity quality.
assignment: assignmenet[i] gives the cluster assignment of node i. 0 <= assignment[i] < ncluster.
ncluster: number of clusters
mq_in: the part of MQ to do with intra-cluster edges, before divide by 1/k
mq_out: the part of MQ to do with inter-cluster edges, before divide by 1/(k*(k-1))
mq = 2*(mq_in/k - mq_out/(k*(k-1)));
*/
int ncluster = 0;
int n = A->m;
int test_pattern_symmetry_only = FALSE;
int *counts, *ia = A->ia, *ja = A->ja, k, i, j, jj;
real mq_in = 0, mq_out = 0, *a = NULL, Vi, Vj;
int c;
real *dout;
assert(SparseMatrix_is_symmetric(A, test_pattern_symmetry_only));
assert(A->n == n);
if (A->type == MATRIX_TYPE_REAL) a = (real*) A->a;
counts = MALLOC(sizeof(int)*n);
for (i = 0; i < n; i++) counts[i] = 0;
for (i = 0; i < n; i++) {
assert(assignment[i] >= 0 && assignment[i] < n);
if (counts[assignment[i]] == 0) ncluster++;
counts[assignment[i]]++;
}
k = ncluster;
assert(ncluster <= n);
for (i = 0; i < n; i++) {
assert(assignment[i] < ncluster);
c = assignment[i];
Vi = counts[c];
for (j = ia[i] ; j < ia[i+1]; j++) {
/* ASSUME UNDIRECTED */
jj = ja[j];
if (jj >= i) continue;
assert(assignment[jj] < ncluster);
Vj = counts[assignment[jj]];
if (assignment[jj] == c) {
if (a) {
mq_in += a[j]/(Vi*Vi);
} else {
mq_in += 1./(Vi*Vi);
}
} else {
if (a) {
mq_out += a[j]/(Vi*Vj);
} else {
mq_out += 1./(Vi*Vj);
}
}
}
}
/* calculate scaled out degree */
dout = MALLOC(sizeof(real)*n);
for (i = 0; i < n; i++) {
dout[i] = 0;
for (j = ia[i]; j < ia[i+1]; j++) {
jj = ja[j];
if (jj == i) continue;
if (a) {
dout[i] += a[j]/(real) counts[assignment[jj]];
} else {
dout[i] += 1./(real) counts[assignment[jj]];
}
}
}
*ncluster0 = k;
*mq_in0 = mq_in;
*mq_out0 = mq_out;
*dout0 = dout;
FREE(counts);
if (k > 1) {
return 2*(mq_in/k - mq_out/(k*(k-1)));
} else {
return 2*mq_in;
}
}
StressMajorizationSmoother SparseStressMajorizationSmoother_new(SparseMatrix A, int dim, real lambda0, real *x,
int weighting_scheme, int scale_initial_coord){
/* solve a stress model to achieve the ideal distance among a sparse set of edges recorded in A.
A must be a real matrix.
*/
StressMajorizationSmoother sm;
int i, j, k, m = A->m, *ia, *ja, *iw, *jw, *id, *jd;
int nz;
real *d, *w, *lambda;
real diag_d, diag_w, *a, dist, s = 0, stop = 0, sbot = 0;
real xdot = 0;
assert(SparseMatrix_is_symmetric(A, FALSE) && A->type == MATRIX_TYPE_REAL);
/* if x is all zero, make it random */
for (i = 0; i < m*dim; i++) xdot += x[i]*x[i];
if (xdot == 0){
for (i = 0; i < m*dim; i++) x[i] = 72*drand();
}
ia = A->ia;
ja = A->ja;
a = (real*) A->a;
sm = MALLOC(sizeof(struct StressMajorizationSmoother_struct));
sm->scaling = 1.;
sm->data = NULL;
sm->scheme = SM_SCHEME_NORMAL;
sm->D = A;
sm->tol_cg = 0.01;
sm->maxit_cg = sqrt((double) A->m);
lambda = sm->lambda = MALLOC(sizeof(real)*m);
for (i = 0; i < m; i++) sm->lambda[i] = lambda0;
nz = A->nz;
sm->Lw = SparseMatrix_new(m, m, nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
sm->Lwd = SparseMatrix_new(m, m, nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
if (!(sm->Lw) || !(sm->Lwd)) {
StressMajorizationSmoother_delete(sm);
return NULL;
}
iw = sm->Lw->ia; jw = sm->Lw->ja;
id = sm->Lwd->ia; jd = sm->Lwd->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
iw[0] = id[0] = 0;
nz = 0;
for (i = 0; i < m; i++){
diag_d = diag_w = 0;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (k != i){
jw[nz] = k;
dist = a[j];
switch (weighting_scheme){
case WEIGHTING_SCHEME_SQR_DIST:
if (dist*dist == 0){
w[nz] = -100000;
} else {
w[nz] = -1/(dist*dist);
}
break;
case WEIGHTING_SCHEME_INV_DIST:
if (dist*dist == 0){
w[nz] = -100000;
} else {
w[nz] = -1/(dist);
}
break;
case WEIGHTING_SCHEME_NONE:
w[nz] = -1;
break;
default:
assert(0);
return NULL;
}
diag_w += w[nz];
jd[nz] = k;
d[nz] = w[nz]*dist;
stop += d[nz]*distance(x,dim,i,k);
sbot += d[nz]*dist;
diag_d += d[nz];
nz++;
}
}
jw[nz] = i;
lambda[i] *= (-diag_w);/* alternatively don't do that then we have a constant penalty term scaled by lambda0 */
w[nz] = -diag_w + lambda[i];
jd[nz] = i;
d[nz] = -diag_d;
nz++;
iw[i+1] = nz;
id[i+1] = nz;
}
if (scale_initial_coord){
s = stop/sbot;
} else {
s = 1.;
}
if (s == 0) {
return NULL;
}
for (i = 0; i < nz; i++) d[i] *= s;
sm->scaling = s;
sm->Lw->nz = nz;
sm->Lwd->nz = nz;
return sm;
}
TriangleSmoother TriangleSmoother_new(SparseMatrix A, int dim, real lambda0, real *x, int use_triangularization){
TriangleSmoother sm;
int i, j, k, m = A->m, *ia = A->ia, *ja = A->ja, *iw, *jw, jdiag, nz;
SparseMatrix B;
real *avg_dist, *lambda, *d, *w, diag_d, diag_w, dist;
real s = 0, stop = 0, sbot = 0;
assert(SparseMatrix_is_symmetric(A, FALSE));
avg_dist = N_GNEW(m,real);
for (i = 0; i < m ;i++){
avg_dist[i] = 0;
nz = 0;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
avg_dist[i] += distance(x, dim, i, ja[j]);
nz++;
}
assert(nz > 0);
avg_dist[i] /= nz;
}
sm = N_GNEW(1,struct TriangleSmoother_struct);
sm->scaling = 1;
sm->data = NULL;
sm->scheme = SM_SCHEME_NORMAL;
sm->tol_cg = 0.01;
sm->maxit_cg = sqrt((double) A->m);
lambda = sm->lambda = N_GNEW(m,real);
for (i = 0; i < m; i++) sm->lambda[i] = lambda0;
if (m > 2){
if (use_triangularization){
B= call_tri(m, dim, x);
} else {
B= call_tri2(m, dim, x);
}
} else {
B = SparseMatrix_copy(A);
}
sm->Lw = SparseMatrix_add(A, B);
SparseMatrix_delete(B);
sm->Lwd = SparseMatrix_copy(sm->Lw);
if (!(sm->Lw) || !(sm->Lwd)) {
TriangleSmoother_delete(sm);
return NULL;
}
iw = sm->Lw->ia; jw = sm->Lw->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
for (i = 0; i < m; i++){
diag_d = diag_w = 0;
jdiag = -1;
for (j = iw[i]; j < iw[i+1]; j++){
k = jw[j];
if (k == i){
jdiag = j;
continue;
}
/* w[j] = -1./(ia[i+1]-ia[i]+ia[ja[j]+1]-ia[ja[j]]);
w[j] = -2./(avg_dist[i]+avg_dist[k]);
w[j] = -1.*/;/* use unit weight for now, later can try 1/(deg(i)+deg(k)) */
dist = pow(distance_cropped(x,dim,i,k),0.6);
w[j] = 1/(dist*dist);
diag_w += w[j];
/* d[j] = w[j]*distance(x,dim,i,k);
d[j] = w[j]*(avg_dist[i] + avg_dist[k])*0.5;*/
d[j] = w[j]*dist;
stop += d[j]*distance(x,dim,i,k);
sbot += d[j]*dist;
diag_d += d[j];
}
lambda[i] *= (-diag_w);/* alternatively don't do that then we have a constant penalty term scaled by lambda0 */
assert(jdiag >= 0);
w[jdiag] = -diag_w + lambda[i];
d[jdiag] = -diag_d;
}
s = stop/sbot;
for (i = 0; i < iw[m]; i++) d[i] *= s;
sm->scaling = s;
FREE(avg_dist);
return sm;
}
static void maximal_independent_edge_set_heavest_edge_pernode_leaves_first(SparseMatrix A, int randomize, int **cluster, int **clusterp, int *ncluster){
int i, ii, j, *ia, *ja, m, n, *p = NULL, q;
real *a, amax = 0;
int first = TRUE, jamax = 0;
int *matched, nz, ncmax = 0, nz0, nzz,k ;
enum {UNMATCHED = -2, MATCHED = -1};
assert(A);
assert(SparseMatrix_known_strucural_symmetric(A));
ia = A->ia;
ja = A->ja;
m = A->m;
n = A->n;
assert(n == m);
*cluster = N_GNEW(m,int);
*clusterp = N_GNEW((m+1),int);
matched = N_GNEW(m,int);
for (i = 0; i < m; i++) matched[i] = i;
assert(SparseMatrix_is_symmetric(A, FALSE));
assert(A->type == MATRIX_TYPE_REAL);
*ncluster = 0;
(*clusterp)[0] = 0;
nz = 0;
a = (real*) A->a;
if (!randomize){
for (i = 0; i < m; i++){
if (matched[i] == MATCHED || node_degree(i) != 1) continue;
q = ja[ia[i]];
assert(matched[q] != MATCHED);
matched[q] = MATCHED;
(*cluster)[nz++] = q;
for (j = ia[q]; j < ia[q+1]; j++){
if (q == ja[j]) continue;
if (node_degree(ja[j]) == 1){
matched[ja[j]] = MATCHED;
(*cluster)[nz++] = ja[j];
}
}
ncmax = MAX(ncmax, nz - (*clusterp)[*ncluster]);
nz0 = (*clusterp)[*ncluster];
if (nz - nz0 <= MAX_CLUSTER_SIZE){
(*clusterp)[++(*ncluster)] = nz;
} else {
(*clusterp)[++(*ncluster)] = ++nz0;
nzz = nz0;
for (k = nz0; k < nz && nzz < nz; k++){
nzz += MAX_CLUSTER_SIZE - 1;
nzz = MIN(nz, nzz);
(*clusterp)[++(*ncluster)] = nzz;
}
}
}
#ifdef DEBUG_print
if (Verbose)
fprintf(stderr, "%d leaves and parents for %d clusters, largest cluster = %d\n",nz, *ncluster, ncmax);
#endif
for (i = 0; i < m; i++){
first = TRUE;
if (matched[i] == MATCHED) continue;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if (matched[ja[j]] != MATCHED && matched[i] != MATCHED){
if (first) {
amax = a[j];
jamax = ja[j];
first = FALSE;
} else {
if (a[j] > amax){
amax = a[j];
jamax = ja[j];
}
}
}
}
if (!first){
matched[jamax] = MATCHED;
matched[i] = MATCHED;
(*cluster)[nz++] = i;
(*cluster)[nz++] = jamax;
(*clusterp)[++(*ncluster)] = nz;
}
}
/* dan yi dian, wu ban */
for (i = 0; i < m; i++){
if (matched[i] == i){
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
assert(nz == n);
} else {
p = random_permutation(m);
for (ii = 0; ii < m; ii++){
i = p[ii];
if (matched[i] == MATCHED || node_degree(i) != 1) continue;
q = ja[ia[i]];
assert(matched[q] != MATCHED);
matched[q] = MATCHED;
(*cluster)[nz++] = q;
for (j = ia[q]; j < ia[q+1]; j++){
if (q == ja[j]) continue;
if (node_degree(ja[j]) == 1){
matched[ja[j]] = MATCHED;
(*cluster)[nz++] = ja[j];
}
}
ncmax = MAX(ncmax, nz - (*clusterp)[*ncluster]);
nz0 = (*clusterp)[*ncluster];
if (nz - nz0 <= MAX_CLUSTER_SIZE){
(*clusterp)[++(*ncluster)] = nz;
} else {
(*clusterp)[++(*ncluster)] = ++nz0;
nzz = nz0;
for (k = nz0; k < nz && nzz < nz; k++){
nzz += MAX_CLUSTER_SIZE - 1;
nzz = MIN(nz, nzz);
(*clusterp)[++(*ncluster)] = nzz;
}
}
}
#ifdef DEBUG_print
if (Verbose)
fprintf(stderr, "%d leaves and parents for %d clusters, largest cluster = %d\n",nz, *ncluster, ncmax);
#endif
for (ii = 0; ii < m; ii++){
i = p[ii];
first = TRUE;
if (matched[i] == MATCHED) continue;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if (matched[ja[j]] != MATCHED && matched[i] != MATCHED){
if (first) {
amax = a[j];
jamax = ja[j];
first = FALSE;
} else {
if (a[j] > amax){
amax = a[j];
jamax = ja[j];
}
}
}
}
if (!first){
matched[jamax] = MATCHED;
matched[i] = MATCHED;
(*cluster)[nz++] = i;
(*cluster)[nz++] = jamax;
(*clusterp)[++(*ncluster)] = nz;
}
}
/* dan yi dian, wu ban */
for (i = 0; i < m; i++){
if (matched[i] == i){
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
FREE(p);
}
FREE(matched);
}
static void maximal_independent_edge_set_heavest_cluster_pernode_leaves_first(SparseMatrix A, int csize,
int randomize, int **cluster, int **clusterp, int *ncluster){
int i, ii, j, *ia, *ja, m, n, *p = NULL, q, iv;
real *a;
int *matched, nz, nz0, nzz,k, nv;
enum {UNMATCHED = -2, MATCHED = -1};
real *vlist;
assert(A);
assert(SparseMatrix_known_strucural_symmetric(A));
ia = A->ia;
ja = A->ja;
m = A->m;
n = A->n;
assert(n == m);
*cluster = N_GNEW(m,int);
*clusterp = N_GNEW((m+1),int);
matched = N_GNEW(m,int);
vlist = N_GNEW(2*m,real);
for (i = 0; i < m; i++) matched[i] = i;
assert(SparseMatrix_is_symmetric(A, FALSE));
assert(A->type == MATRIX_TYPE_REAL);
*ncluster = 0;
(*clusterp)[0] = 0;
nz = 0;
a = (real*) A->a;
p = random_permutation(m);
for (ii = 0; ii < m; ii++){
i = p[ii];
if (matched[i] == MATCHED || node_degree(i) != 1) continue;
q = ja[ia[i]];
assert(matched[q] != MATCHED);
matched[q] = MATCHED;
(*cluster)[nz++] = q;
for (j = ia[q]; j < ia[q+1]; j++){
if (q == ja[j]) continue;
if (node_degree(ja[j]) == 1){
matched[ja[j]] = MATCHED;
(*cluster)[nz++] = ja[j];
}
}
nz0 = (*clusterp)[*ncluster];
if (nz - nz0 <= MAX_CLUSTER_SIZE){
(*clusterp)[++(*ncluster)] = nz;
} else {
(*clusterp)[++(*ncluster)] = ++nz0;
nzz = nz0;
for (k = nz0; k < nz && nzz < nz; k++){
nzz += MAX_CLUSTER_SIZE - 1;
nzz = MIN(nz, nzz);
(*clusterp)[++(*ncluster)] = nzz;
}
}
}
for (ii = 0; ii < m; ii++){
i = p[ii];
if (matched[i] == MATCHED) continue;
nv = 0;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if (matched[ja[j]] != MATCHED && matched[i] != MATCHED){
vlist[2*nv] = ja[j];
vlist[2*nv+1] = a[j];
nv++;
}
}
if (nv > 0){
qsort(vlist, nv, sizeof(real)*2, scomp);
for (j = 0; j < MIN(csize - 1, nv); j++){
iv = (int) vlist[2*j];
matched[iv] = MATCHED;
(*cluster)[nz++] = iv;
}
matched[i] = MATCHED;
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
/* dan yi dian, wu ban */
for (i = 0; i < m; i++){
if (matched[i] == i){
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
FREE(p);
FREE(matched);
}
static void maximal_independent_edge_set_heavest_edge_pernode_supernodes_first(SparseMatrix A, int randomize, int **cluster, int **clusterp, int *ncluster){
int i, ii, j, *ia, *ja, m, n, *p = NULL;
real *a, amax = 0;
int first = TRUE, jamax = 0;
int *matched, nz, nz0;
enum {UNMATCHED = -2, MATCHED = -1};
int nsuper, *super = NULL, *superp = NULL;
assert(A);
assert(SparseMatrix_known_strucural_symmetric(A));
ia = A->ia;
ja = A->ja;
m = A->m;
n = A->n;
assert(n == m);
*cluster = N_GNEW(m,int);
*clusterp = N_GNEW((m+1),int);
matched = N_GNEW(m,int);
for (i = 0; i < m; i++) matched[i] = i;
assert(SparseMatrix_is_symmetric(A, FALSE));
assert(A->type == MATRIX_TYPE_REAL);
SparseMatrix_decompose_to_supervariables(A, &nsuper, &super, &superp);
*ncluster = 0;
(*clusterp)[0] = 0;
nz = 0;
a = (real*) A->a;
for (i = 0; i < nsuper; i++){
if (superp[i+1] - superp[i] <= 1) continue;
nz0 = (*clusterp)[*ncluster];
for (j = superp[i]; j < superp[i+1]; j++){
matched[super[j]] = MATCHED;
(*cluster)[nz++] = super[j];
if (nz - nz0 >= MAX_CLUSTER_SIZE){
(*clusterp)[++(*ncluster)] = nz;
nz0 = nz;
}
}
if (nz > nz0) (*clusterp)[++(*ncluster)] = nz;
}
if (!randomize){
for (i = 0; i < m; i++){
first = TRUE;
if (matched[i] == MATCHED) continue;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if (matched[ja[j]] != MATCHED && matched[i] != MATCHED){
if (first) {
amax = a[j];
jamax = ja[j];
first = FALSE;
} else {
if (a[j] > amax){
amax = a[j];
jamax = ja[j];
}
}
}
}
if (!first){
matched[jamax] = MATCHED;
matched[i] = MATCHED;
(*cluster)[nz++] = i;
(*cluster)[nz++] = jamax;
(*clusterp)[++(*ncluster)] = nz;
}
}
/* dan yi dian, wu ban */
for (i = 0; i < m; i++){
if (matched[i] == i){
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
assert(nz == n);
} else {
p = random_permutation(m);
for (ii = 0; ii < m; ii++){
i = p[ii];
first = TRUE;
if (matched[i] == MATCHED) continue;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if (matched[ja[j]] != MATCHED && matched[i] != MATCHED){
if (first) {
amax = a[j];
jamax = ja[j];
first = FALSE;
} else {
if (a[j] > amax){
amax = a[j];
jamax = ja[j];
}
}
}
}
if (!first){
matched[jamax] = MATCHED;
matched[i] = MATCHED;
(*cluster)[nz++] = i;
(*cluster)[nz++] = jamax;
(*clusterp)[++(*ncluster)] = nz;
}
}
/* dan yi dian, wu ban */
for (i = 0; i < m; i++){
if (matched[i] == i){
(*cluster)[nz++] = i;
(*clusterp)[++(*ncluster)] = nz;
}
}
FREE(p);
}
FREE(super);
FREE(superp);
FREE(matched);
}
SparseMatrix ideal_distance_matrix(SparseMatrix A, int dim, real *x){
/* find the ideal distance between edges, either 1, or |N[i] \Union N[j]| - |N[i] \Intersection N[j]|
*/
SparseMatrix D;
int *ia, *ja, i, j, k, l, nz;
real *d;
int *mask = NULL;
real len, di, sum, sumd;
assert(SparseMatrix_is_symmetric(A, FALSE));
D = SparseMatrix_copy(A);
ia = D->ia;
ja = D->ja;
if (D->type != MATRIX_TYPE_REAL){
FREE(D->a);
D->type = MATRIX_TYPE_REAL;
D->a = N_GNEW(D->nz,real);
}
d = (real*) D->a;
mask = N_GNEW(D->m,int);
for (i = 0; i < D->m; i++) mask[i] = -1;
for (i = 0; i < D->m; i++){
di = node_degree(i);
mask[i] = i;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
mask[ja[j]] = i;
}
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (i == k) continue;
len = di + node_degree(k);
for (l = ia[k]; l < ia[k+1]; l++){
if (mask[ja[l]] == i) len--;
}
d[j] = len;
assert(len > 0);
}
}
sum = 0; sumd = 0;
nz = 0;
for (i = 0; i < D->m; i++){
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
nz++;
sum += distance(x, dim, i, ja[j]);
sumd += d[j];
}
}
sum /= nz; sumd /= nz;
sum = sum/sumd;
for (i = 0; i < D->m; i++){
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
d[j] = sum*d[j];
}
}
return D;
}
static void maximal_independent_edge_set_heavest_edge_pernode_scaled(SparseMatrix A, int randomize, int **matching, int *nmatch){
int i, ii, j, *ia, *ja, m, n, *p = NULL;
real *a, amax = 0;
int first = TRUE, jamax = 0;
assert(A);
assert(SparseMatrix_known_strucural_symmetric(A));
ia = A->ia;
ja = A->ja;
m = A->m;
n = A->n;
assert(n == m);
*matching = N_GNEW(m,int);
for (i = 0; i < m; i++) (*matching)[i] = i;
*nmatch = n;
assert(SparseMatrix_is_symmetric(A, FALSE));
assert(A->type == MATRIX_TYPE_REAL);
a = (real*) A->a;
if (!randomize){
for (i = 0; i < m; i++){
first = TRUE;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if ((*matching)[ja[j]] == ja[j] && (*matching)[i] == i){
if (first) {
amax = a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]);
jamax = ja[j];
first = FALSE;
} else {
if (a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]) > amax){
amax = a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]);
jamax = ja[j];
}
}
}
}
if (!first){
(*matching)[jamax] = i;
(*matching)[i] = jamax;
(*nmatch)--;
}
}
} else {
p = random_permutation(m);
for (ii = 0; ii < m; ii++){
i = p[ii];
if ((*matching)[i] != i) continue;
first = TRUE;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
if ((*matching)[ja[j]] == ja[j] && (*matching)[i] == i){
if (first) {
amax = a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]);
jamax = ja[j];
first = FALSE;
} else {
if (a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]) > amax){
amax = a[j]/(ia[i+1]-ia[i])/(ia[ja[j]+1]-ia[ja[j]]);
jamax = ja[j];
}
}
}
}
if (!first){
(*matching)[jamax] = i;
(*matching)[i] = jamax;
(*nmatch)--;
}
}
FREE(p);
}
}
StressMajorizationSmoother StressMajorizationSmoother2_new(SparseMatrix A, int dim, real lambda0, real *x,
int ideal_dist_scheme){
/* use up to dist 2 neighbor */
/* use up to dist 2 neighbor. This is used in overcoming pherical effect with ideal distance of
2-neighbors equal graph distance etc.
*/
StressMajorizationSmoother sm;
int i, j, k, l, m = A->m, *ia = A->ia, *ja = A->ja, *iw, *jw, *id, *jd;
int *mask, nz;
real *d, *w, *lambda;
real *avg_dist, diag_d, diag_w, dist, s = 0, stop = 0, sbot = 0;
SparseMatrix ID;
assert(SparseMatrix_is_symmetric(A, FALSE));
ID = ideal_distance_matrix(A, dim, x);
sm = GNEW(struct StressMajorizationSmoother_struct);
sm->scaling = 1.;
sm->data = NULL;
sm->scheme = SM_SCHEME_NORMAL;
sm->tol_cg = 0.01;
sm->maxit_cg = sqrt((double) A->m);
lambda = sm->lambda = N_GNEW(m,real);
for (i = 0; i < m; i++) sm->lambda[i] = lambda0;
mask = N_GNEW(m,int);
avg_dist = N_GNEW(m,real);
for (i = 0; i < m ;i++){
avg_dist[i] = 0;
nz = 0;
for (j = ia[i]; j < ia[i+1]; j++){
if (i == ja[j]) continue;
avg_dist[i] += distance(x, dim, i, ja[j]);
nz++;
}
assert(nz > 0);
avg_dist[i] /= nz;
}
for (i = 0; i < m; i++) mask[i] = -1;
nz = 0;
for (i = 0; i < m; i++){
mask[i] = i;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (mask[k] != i){
mask[k] = i;
nz++;
}
}
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
for (l = ia[k]; l < ia[k+1]; l++){
if (mask[ja[l]] != i){
mask[ja[l]] = i;
nz++;
}
}
}
}
sm->Lw = SparseMatrix_new(m, m, nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
sm->Lwd = SparseMatrix_new(m, m, nz + m, MATRIX_TYPE_REAL, FORMAT_CSR);
if (!(sm->Lw) || !(sm->Lwd)) {
StressMajorizationSmoother_delete(sm);
return NULL;
}
iw = sm->Lw->ia; jw = sm->Lw->ja;
w = (real*) sm->Lw->a; d = (real*) sm->Lwd->a;
id = sm->Lwd->ia; jd = sm->Lwd->ja;
iw[0] = id[0] = 0;
nz = 0;
for (i = 0; i < m; i++){
mask[i] = i+m;
diag_d = diag_w = 0;
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
if (mask[k] != i+m){
mask[k] = i+m;
jw[nz] = k;
if (ideal_dist_scheme == IDEAL_GRAPH_DIST){
dist = 1;
} else if (ideal_dist_scheme == IDEAL_AVG_DIST){
dist = (avg_dist[i] + avg_dist[k])*0.5;
} else if (ideal_dist_scheme == IDEAL_POWER_DIST){
dist = pow(distance_cropped(x,dim,i,k),.4);
} else {
fprintf(stderr,"ideal_dist_scheme value wrong");
assert(0);
exit(1);
}
/*
w[nz] = -1./(ia[i+1]-ia[i]+ia[ja[j]+1]-ia[ja[j]]);
w[nz] = -2./(avg_dist[i]+avg_dist[k]);*/
/* w[nz] = -1.;*//* use unit weight for now, later can try 1/(deg(i)+deg(k)) */
w[nz] = -1/(dist*dist);
diag_w += w[nz];
jd[nz] = k;
/*
d[nz] = w[nz]*distance(x,dim,i,k);
d[nz] = w[nz]*dd[j];
d[nz] = w[nz]*(avg_dist[i] + avg_dist[k])*0.5;
*/
d[nz] = w[nz]*dist;
stop += d[nz]*distance(x,dim,i,k);
sbot += d[nz]*dist;
diag_d += d[nz];
nz++;
}
}
/* distance 2 neighbors */
for (j = ia[i]; j < ia[i+1]; j++){
k = ja[j];
for (l = ia[k]; l < ia[k+1]; l++){
if (mask[ja[l]] != i+m){
mask[ja[l]] = i+m;
if (ideal_dist_scheme == IDEAL_GRAPH_DIST){
dist = 2;
} else if (ideal_dist_scheme == IDEAL_AVG_DIST){
dist = (avg_dist[i] + 2*avg_dist[k] + avg_dist[ja[l]])*0.5;
} else if (ideal_dist_scheme == IDEAL_POWER_DIST){
dist = pow(distance_cropped(x,dim,i,ja[l]),.4);
} else {
fprintf(stderr,"ideal_dist_scheme value wrong");
assert(0);
exit(1);
}
jw[nz] = ja[l];
/*
w[nz] = -1/(ia[i+1]-ia[i]+ia[ja[l]+1]-ia[ja[l]]);
w[nz] = -2/(avg_dist[i] + 2*avg_dist[k] + avg_dist[ja[l]]);*/
/* w[nz] = -1.;*//* use unit weight for now, later can try 1/(deg(i)+deg(k)) */
w[nz] = -1/(dist*dist);
diag_w += w[nz];
jd[nz] = ja[l];
/*
d[nz] = w[nz]*(distance(x,dim,i,k)+distance(x,dim,k,ja[l]));
d[nz] = w[nz]*(dd[j]+dd[l]);
d[nz] = w[nz]*(avg_dist[i] + 2*avg_dist[k] + avg_dist[ja[l]])*0.5;
*/
d[nz] = w[nz]*dist;
stop += d[nz]*distance(x,dim,ja[l],k);
sbot += d[nz]*dist;
diag_d += d[nz];
nz++;
}
}
}
jw[nz] = i;
lambda[i] *= (-diag_w);/* alternatively don't do that then we have a constant penalty term scaled by lambda0 */
w[nz] = -diag_w + lambda[i];
jd[nz] = i;
d[nz] = -diag_d;
nz++;
iw[i+1] = nz;
id[i+1] = nz;
}
s = stop/sbot;
for (i = 0; i < nz; i++) d[i] *= s;
sm->scaling = s;
sm->Lw->nz = nz;
sm->Lwd->nz = nz;
FREE(mask);
FREE(avg_dist);
SparseMatrix_delete(ID);
return sm;
}
