void testCoarsening(int *nvtxs, idxtype *xadj, idxtype *adjncy,
idxtype *vwgt, idxtype *adjwgt,
int *wgtflag, int ct)
{
GraphType graph, *cgraph;
CtrlType ctrl;
my_SetUpGraph(&graph, *nvtxs, xadj, adjncy, vwgt, adjwgt,
*wgtflag, 1);
/* The last argument indicates we are setting up the original
* graph */
ctrl.CoarsenTo=ct;
ctrl.CType=MATCH_SHEMN;
my_AllocateWorkSpace(&ctrl,&graph);
cgraph = Coarsen2Way(&ctrl,&graph);
do {
dump_graph(cgraph);
cgraph = cgraph->finer;
} while ( cgraph != NULL );
}
void srmcl(int* nvtxs, idxtype* xadj, idxtype* adjncy, idxtype
*vwgt, idxtype* adjwgt, int* wgtflag, std::vector<std::set<idxtype> >& indices, Options opt)
{
GraphType *graph = (GraphType*)malloc(sizeof(GraphType));
my_SetUpGraph(graph, *nvtxs, xadj, adjncy, vwgt, adjwgt, *wgtflag, 1);
std::vector< std::set<idxtype> > clusters = srmclWithGraph(graph, indices, opt); //main procedure!
//post processing: prune out clusters
printf("start post-processing - prune clusters.\n");
int num_clusters = clusters.size();
double max_weight = 10000;
//printf("double max_weight = 10000;\n");
//prune out clusters according to their clustering coefficient
#ifdef CLUSTER_COEFFICIENT
double* cluster_coefficients = (double*) malloc(sizeof(double) * num_clusters);
//weighted cc is according to [B. Zhang and S. Horvath, Stat. App. Genet. Mol. Biol. 4, 17 2005.]
for(int vIdx = 0; vIdx < *nvtxs; vIdx++){ //v1
for(std::set<idxtype>::iterator cID = indices[vIdx].begin(); cID != indices[vIdx].end(); cID ++){ //for each cluster, calculate its clustering coefficient
double numerator = 0;
double denominator1 = 0;
double denominator2 = 0;
for(int adjIdx = xadj[vIdx]; adjIdx < xadj[vIdx+1]; adjIdx++){
if(indices[adjncy[adjIdx]].find(*cID) != indices[adjncy[adjIdx]].end() && adjncy[adjIdx] != vIdx){ //if another node v1 is also in this cluster
double wij = (double)adjwgt[adjIdx] / max_weight;
denominator1 += wij;
denominator2 += pow(wij,2.0);
for(int adjIdx2 = adjIdx+1; adjIdx2 < xadj[vIdx+1]; adjIdx2++){
if(indices[adjncy[adjIdx2]].find(*cID) != indices[adjncy[adjIdx2]].end() && adjncy[adjIdx2] != vIdx){ //if another node v2 is also in this cluster
double wik = (double)adjwgt[adjIdx2] / max_weight;
for(int i=xadj[adjncy[adjIdx]]; i<xadj[adjncy[adjIdx]+1]; i++){ //find whether v1 and v2 are connected
if(adjncy[i] == adjncy[adjIdx2]){ //if v1 and v2 are connected
double wjk = (double)adjwgt[i] / max_weight;
numerator += wij*wik*wjk;
//if(*cID == 3)
break;
}
}
}
}
}
}
denominator1 = pow(denominator1, 2.0);
double denominator = denominator1 - denominator2;
numerator *= 2;
double cluster_coefficient = 0;
if(numerator != 0)
cluster_coefficient = numerator / denominator;
//printf("cID:%d node:%d value:%.3f (%.3f/%.3f)\n",*cID, vIdx+1,cluster_coefficient,numerator,denominator );
cluster_coefficients[*cID] += cluster_coefficient;
}
}
#endif
//calculate weighted density
double* num_internal_edges = (double*) malloc( sizeof(double) * num_clusters);
for(int i=0; i<num_clusters; i++)
num_internal_edges[i] = 0;
for(int vID=0; vID<*nvtxs ; vID++){
for(std::set<idxtype>::iterator cIterator = indices[vID].begin(); cIterator != indices[vID].end(); cIterator++){
idxtype cID = *(cIterator);
for( int j = xadj[vID]; j < xadj[vID+1]; j++ ){
if( adjncy[j]!=vID && clusters[cID].find( adjncy[j] ) != clusters[cID].end() ){ //contains it
if(opt.weighted_density)
num_internal_edges[ cID ] += (adjwgt[j] / max_weight); //weighted version
else
num_internal_edges[ cID ] ++;
}
}
}
}
//printf("double* densities = (double*) malloc(sizeof(double) * (*nvtxs)); \n"); fflush(stdout);
double* densities = (double*) malloc(sizeof(double) * num_clusters);
for(int cID = 0; cID < num_clusters ; cID++){
int size = clusters[cID].size();
if(size <= 1)
densities[cID] = 0;
else
densities[cID] = num_internal_edges[ cID ] / size / (size-1);
}
free(num_internal_edges);
int num_pruned_clusters_density = 0;
for(int cID = 0; cID < num_clusters ; cID++){
#ifdef CLUSTER_COEFFICIENT
if(clusters[cID].size() > 1){
cluster_coefficients[cID] /= clusters[cID].size();
densities[cID] = cluster_coefficients[cID] ;
}
#endif
#ifdef TEST_OUTPUT
printf("cluster %d: \tdensity:%.3f\tsize:%zu\n",cID, densities[cID], clusters[cID].size());
#endif
if(clusters[cID].size()<=2 || densities[cID] * sqrt((double)clusters[cID].size()) < opt.quality_threshold){ //remove the cluster
//if(clusters[cID].size()<=2 || densities[cID] < opt.quality_threshold){ //remove the cluster
for(std::set<idxtype>::iterator nodeIterator = clusters[cID].begin(); nodeIterator != clusters[cID].end(); nodeIterator++){
//printf(" nodeIterator:%d\n",*nodeIterator);
if(indices[*nodeIterator].find(cID) != indices[*nodeIterator].end()){
//printf(" if nodeIterator:%d\n",*(indices[*nodeIterator].find(cID)));
indices[*nodeIterator].erase(cID);
}
//printf(" nodeIterator:%d done\n",*nodeIterator);
}
num_pruned_clusters_density++;
clusters[cID].clear();
densities[cID] = -1;
#ifdef CLUSTER_COEFFICIENT
cluster_coefficients [cID] = -1;
#endif
}
}
printf("number of clusters pruned out since their density are smaller than %.3f:\t%d\n", opt.quality_threshold, num_pruned_clusters_density);
//prune out clusters according to their redundancy (sort clusters by density * sqrt(size))
int num_pruned_clusters_overlap = 0;
int num_clusters_after_pruning_density = (num_clusters-num_pruned_clusters_density);
printf("sort clusters\n");
//printf("malloc test: %d, sizeof(Cluster):%d (int:%d)\n",num_clusters_after_pruning_density,sizeof(Cluster),sizeof(int)); fflush(stdout);
Cluster* cluster_array = (Cluster*) malloc(sizeof(Cluster)*(num_clusters_after_pruning_density ));
//printf("malloc test2: %d\n",num_clusters_after_pruning_density); fflush(stdout);
int tempIdx = 0;
for(int i=0; i < num_clusters; i++){
if(densities[i] >= 0){
Cluster c(i,clusters[i].size(), densities[i] );
cluster_array[tempIdx] = c;
tempIdx++;
}
}
double avg_cluster_size = 0;
qsort(cluster_array, num_clusters_after_pruning_density , sizeof(Cluster), compareCluster);
for(int i=0; i<num_clusters_after_pruning_density ; i++){
int cID1 = cluster_array[i].cID;
if(clusters[cID1].size() == 0)
continue;
//printf("examined cluster cID1: %d\n",cID1);fflush(stdout);
for(int j=i+1; j<num_clusters_after_pruning_density ; j++){
//calculate overlap size
int cID2 = cluster_array[j].cID;
//printf(" examined cluster cID2: %d\n",cID2);fflush(stdout);
if(clusters[cID2].size() == 0)
continue;
double overlap = 0;
for ( std::set<idxtype>::iterator iterator = clusters[cID2].begin(); iterator != clusters[cID2].end(); iterator++ ){
if ( clusters[cID1].find(*iterator) != clusters[cID1].end() ){
overlap++;
}
}
//printf("overlap: %1.0f\n",overlap);fflush(stdout);
//calculate neighbor affinity
float overlapNA = pow(overlap,2) / clusters[cID1].size() / clusters[cID2].size();
if(overlapNA >= opt.redundancy_threshold){ //remove the cluster with cID2
for(std::set<idxtype>::iterator nodeIterator = clusters[cID2].begin(); nodeIterator != clusters[cID2].end(); nodeIterator++){
indices[*nodeIterator].erase(cID2);
}
num_pruned_clusters_overlap++;
#ifdef TEST_OUTPUT
printf(" overlapNA:%.3f, (keep cID:%d, remove cID:%d) test size:%1f, %d, %d\n",overlapNA,cID1,overlap, cID2,clusters[cID1].size(),clusters[cID2].size());fflush(stdout);
#endif
clusters[cID2].clear();
densities[cID2] = 0;
}
}
avg_cluster_size += clusters[cID1].size();
}
printf("number of overlapped clusters pruned out since their NA are larger than %.3f:\t%d\n", opt.redundancy_threshold, num_pruned_clusters_overlap);fflush(stdout);
num_clusters -= (num_pruned_clusters_density + num_pruned_clusters_overlap);
printf("*********afer prunning, total # clusters:\t%d**********\n",num_clusters);fflush(stdout);
avg_cluster_size /= num_clusters;
printf("*********average cluster size:\t%.3f**********\n",avg_cluster_size);fflush(stdout);
int coverage = 0;
for(int vID=0; vID<*nvtxs ; vID++){
if(indices[vID].size() > 0)
coverage++;
}
printf("*********coverage:\t%d**********\n",coverage);fflush(stdout);
free(cluster_array);
free(graph);
}
void mlmcl(int* nvtxs, idxtype* xadj, idxtype* adjncy, idxtype
*vwgt, idxtype* adjwgt, int* wgtflag, idxtype* indices, Options opt)
{
/* GraphType graph;
my_SetUpGraph(&graph, *nvtxs, xadj, adjncy, vwgt, adjwgt,
*wgtflag, 1); */
int hubRemoval=opt.hubRemoval, recursiveCluster=0;
float hub_pct = opt.hubPct;
GraphType *graph = (GraphType*)malloc(sizeof(GraphType));
my_SetUpGraph(graph, *nvtxs, xadj, adjncy, vwgt, adjwgt,
*wgtflag, 1);
// The last argument indicates we are setting up the original
// graph
idxtype* newIds;
if ( hubRemoval > 0 )
{
int hubThreshold = (int) floor(hub_pct * graph->nvtxs);
GraphType *new_graph;
newIds = removeHubs(graph, hubThreshold, *wgtflag,
&new_graph, 0);
free(graph->gdata);
free(graph);
graph = new_graph;
// now need to remove any nodes that became singletons
// because of hub removal.
// we'll do another iteration of newIds, so back up
// the old newIds. newIds_bkp is of size *nvtxs.
idxtype *newIds_bkp = newIds;
int noOfSingletons = 0, newIdCounter;
newIds = lookForSingletons(graph, &noOfSingletons);
newIdCounter = graph->nvtxs - noOfSingletons;
if ( noOfSingletons > 0 )
{
printf("%d nodes became singletons due to hub removal",
noOfSingletons );
printf("; they will be removed.\n");
fflush(stdout);
getSubgraph(graph, newIds, newIdCounter, *wgtflag,
&new_graph);
free(graph->gdata);
free(graph);
graph = new_graph;
int i;
for ( i=0; i<*nvtxs; i++ )
{
if ( newIds_bkp[i] > -1 )
{
newIds_bkp[i] = newIds[newIds_bkp[i]];
}
else
newIds_bkp[i] = -1;
}
free(newIds);
}
newIds=newIds_bkp;
}
// printf("nnz:%d\n",graph.xadj[*nvtxs]);
if ( opt.mis_coarsenType > 0 )
{
// mis_mlrmcl(graph, indices, opt);
}
else
{
mlmclWithGraph(graph, indices, opt);
}
if ( hubRemoval > 0 )
{
int npart=mapPartition(indices, graph->nvtxs);
float ncut=ComputeNCut(graph, indices, npart);
printf("In graph that does not include hubs,");
printf("No. of Clusters:%d, N-Cut value: %.2f\n", npart, ncut);
mapIndices(indices, newIds, *nvtxs, npart);
free(newIds);
if ( *nvtxs - graph->nvtxs > 0 )
{
char filename[256];
sprintf(filename, "input.nohubs.%.3f", hub_pct);
WriteGraph(filename, graph->nvtxs, graph->xadj,
graph->adjncy);
printf("Wrote nohubs graph to %s\n", filename);
}
}
if ( recursiveCluster > 0 )
{
int npart = mapPartition(indices, graph->nvtxs);
float ncut = ComputeNCut(graph, indices, npart);
printf("No. of clusters:%d, N-Cut:%.2f\n", npart, ncut);
idxtype* hist = histogram(indices, graph->nvtxs, npart);
int max=0, i=0, maxCluster=-1;
for( i=0; i<npart; i++ )
{
if ( hist[i] > max )
{
max = hist[i];
maxCluster = i;
}
}
free(hist);
if ( max > graph->nvtxs * 0.3 )
{
printf("Will recursively partition cluster of size");
printf(" %d\n", max);
idxtype* newIds = idxmalloc(graph->nvtxs,"mlmcl:newIds");
int newIdCounter=0;
for ( i=0; i<graph->nvtxs; i++ )
{
if ( indices[i] == maxCluster )
newIds[i]=newIdCounter++;
else
newIds[i]=-1;
}
GraphType *new_graph;
getSubgraph(graph, newIds, max, *wgtflag, &new_graph);
idxtype *new_indices = idxmalloc(max,"mlmcl:new_indices");
opt.coarsenTo = (int) round(((float) max
/ (float)graph->nvtxs) * opt.coarsenTo);
mlmcl(&max,new_graph->xadj, new_graph->adjncy,
new_graph->vwgt, new_graph->adjwgt, wgtflag,
new_indices, opt );
int new_npart = mapPartition( new_indices, max);
for ( i=0; i<graph->nvtxs; i++ )
{
if ( newIds[i] > -1 )
{
int ni = new_indices[newIds[i]];
if ( ni > 0 )
indices[newIds[i]] = npart + ni - 1;
else
indices[newIds[i]] = maxCluster;
}
}
printf("Recursive clustering yielded %d new",new_npart);
printf(" clusters.");
free(new_indices);
free(newIds);
free(new_graph->gdata);
free(new_graph);
}
}
}
