void KMeans::run()
{
bool move;
bool some_point_is_moving = true;
unsigned int num_iterations = 0;
PointId pid;
ClusterId cid, to_cluster;
Distance d, min;
// Initial partition of points
initial_partition_points();
// Until not converge
while (some_point_is_moving)
{
some_point_is_moving = false;
compute_centroids();
// for each point
for (pid=0; pid<n_; ++pid)
{
// distance from current cluster
ClusterId curr_ID = points_to_clusters_[pid];
Point curr_centroid = centroids_.row(curr_ID);
min = distance(curr_centroid, X_.getdatai(pid));
// for each centroid
cid = 0;
move = false;
for(int k=0; k<k_;++k)
{
d = distance(centroids_.row(k), X_.getdatai(pid));
if (d < min)
{
min = d;
move = true;
to_cluster = cid;
// remove from current cluster
clusters_to_points_[points_to_clusters_[pid]].erase(pid);
some_point_is_moving = true;
}
cid++;
}
// move towards a closer centroid
if (move)
{
// insert
points_to_clusters_[pid] = to_cluster;
clusters_to_points_[to_cluster].insert(pid);
}
}
num_iterations++;
}
std::cout<<"Num Iter: "<<num_iterations<<std::endl;
}
/*
* Clusters data.
*/
static void kmeans(void)
{
int i;
omp_set_num_threads(NUM_THREADS);
for (i = 0; i < NUM_THREADS; i++)
omp_init_lock(&lock[i]);
/* Cluster data. */
do
{
populate();
compute_centroids();
} while (again());
}
void Clusters::generate(RecTable & mods0,
int nreclusters,
int niterations,
bool grahamSchmidt)
{
bool verbose = true;
cout << "Generating clusters using " << mods0.size() << " users" << endl;
int i;
if(!mods0.size()) {
cout << "Bad: mods table uninitialized" << endl;
return;
}
if(verbose) {
cout << "Mods table has " << mods0.size() << " users" << endl;
cout << "generating clusters" << endl;
}
clear();
// default to every user in a random bin
ModsTable::iterator it;
for(it = mods0.begin(), i=0;
it != mods0.end(); i++, it++) {
if(it == mods0.end()) {
cout << "Bad: mods table is smaller than the number of clusters" << endl;
return;
}
clusters[i % nclusters].push_back((*it).first);
}
cout << "Number of clusters: " << clusters.size() << endl;
// recompute them
compute_centroids();
restart_this_sucker:
if (grahamSchmidt) GrahamSchmidt(1);
for(int iter = 0; iter < niterations; iter++) {
time_t start_time = time(NULL);
if(verbose)
cout << "Iteration: " << iter << endl;
// wipe the existing clusters
clear();
if(verbose)
cout << " --> Binning users" << endl;
int i = 0;
for(ModsTable::iterator user = mods0.begin();
user != mods0.end(); user++, i++)
add((*user).first);
if(verbose)
cout << " --> Copying current centroids" << endl;
// make a copy of the current centroids
vector<User> oldcentroids;
for(i = 0; i < nclusters; i++) {
oldcentroids.push_back(User());
for(User::iterator article = centroids[i].begin();
article != centroids[i].end(); article++) {
oldcentroids[i].add((*article).first,
centroids[i][(*article).first]);
}
}
// recompute them
compute_centroids();
float quality=0, difference=0;
int n_non_zero = 0;
for(i = 0; i < nclusters; i++) {
if(clusters[i].empty()) continue;
n_non_zero++;
float q = spread[i];
quality += q;
// float d = centroids[i] | oldcentroids[i];
float d = dist(¢roids[i], &oldcentroids[i], true);
difference += d;
int n_per_user = 0;
for(vector<int>::iterator it = clusters[i].begin();
it != clusters[i].end(); it++) {
n_per_user += mods0[(*it)].size();
}
if(verbose)
cout << "\t( #" << setw(2) << i
<< ", Q:" << setprecision(3) << setw(7)
<< q << ", Delta:" << setprecision(3) << setw(7) << d
<< ", Users:" << setw(6) << clusters[i].size()
<< ", Arts:" << setw(6) << centroids[i].size()
<< ", Avg # art:" << setw(6) << float(n_per_user)/float(clusters[i].size())
<< ")"
<< endl;
}
if(n_non_zero) {
quality /= float(n_non_zero);
difference /= float(n_non_zero);
}
if(verbose)
cout << "Quality " << quality << ", Difference " << difference
<< ": Total time " << difftime(time(NULL), start_time) << "sec"
<< endl;
if(difference > .98) break;
}
// dump the self-similarity matrix for the centroids
cout << endl << "\x1b[31mCentroid Comparison Matrix\x1b[0m" << endl;
for(i = 0; i < nclusters; i++) {
printf("\t");
for(int j = 0; j < nclusters; j++) {
printf("%6.3f", dist(¢roids[i], ¢roids[j], true));
if(j != nclusters -1)
printf(",");
}
printf("\n");
}
// dump the self-similarity matrix for the centroids
cout << endl << "\x1b[31mCentroid Comparison Matrix (dot)\x1b[0m" << endl;
for(i = 0; i < nclusters; i++) {
printf("\t");
for(int j = 0; j < nclusters; j++) {
printf("%8.3f", dot(¢roids[i], ¢roids[j]));
if(j != nclusters -1)
printf(",");
}
printf("\n");
}
// and the corresponding number of entries per cluster
cout << endl<< "\x1b[31m Article availability\x1b[0m" << endl;
for(i = 0; i < nclusters; i++) {
cout << "\tCluster #" << i << ": " << centroids[i].size()
<< ", Norm: " << norm(¢roids[i]) << endl;
}
if(nreclusters --) goto restart_this_sucker;
}
