void rank_cross_validation(const problem *prob, const parameter *param, int nr_fold, double *result)
{
int i,q;
int *fold_start;
int l = prob->l;
int *query_set;
double *target = Malloc(double,l);
int nr_query;
int *start = NULL;
int *count = NULL;
int *perm = Malloc(int,l);
int *query_perm;
group_queries(prob->query, l, &nr_query, &start, &count, perm);
if (nr_query == 1)
{
if (nr_fold > prob->l / 2)
{
nr_fold = l / 2; // each fold should include at least 2 instances to form pairs
fprintf(stderr,"WARNING: # folds > # data / 2. Will use # folds = # data / 2 instead (Every fold should contain 2 data to form a pair)\n");
}
nr_query = nr_fold;
// Treat each fold as a query in performance evaluation
// to avoid ranking inconsistency between models.
start = (int *)realloc(start,nr_query * sizeof(int));
count = (int *)realloc(count,nr_query * sizeof(int));
query_set = Malloc(int,l);
for(q=0;q<nr_query;q++)
{
count[q] = 0;
start[q] = 0;
}
for(i=0;i<l;i++)
{
int j = rand() % nr_query;
query_set[i] = j;
count[j]++;
}
start[0] = 0;
for(q=1;q<nr_query;q++)
start[q] = start[q-1] + count[q-1];
for(i=0;i<l;i++)
{
perm[start[query_set[i]]] = i;
++start[query_set[i]];
}
start[0] = 0;
for(q=1;q<nr_query;q++)
start[q] = start[q-1] + count[q-1];
}
else
{
l2r_l2_ranksvm_fun::l2r_l2_ranksvm_fun(const problem *prob, double C)
{
int q,i,k;
int l=prob->l;
double *y = prob->y;
this->prob = prob;
z = new double[l];
l_plus = new int[l];
l_minus = new int[l];
alpha_plus = new double[l];
alpha_minus = new double[l];
int_y = new int[prob->l];
this->C = C;
// Checking number of relevance levels in each query
// Meanwhile transform the labels into 1,...,k
perm = Malloc(int,l);
group_queries(prob->query, l, &nr_query, &start, &count, perm);
nr_class = new int[nr_query];
order_perm = new id_and_value *[nr_query];
for (q=0;q<nr_query;q++)
order_perm[q] = new id_and_value[count[q]];
for (q=0;q<nr_query;q++)
{
int *perm_q = &perm[start[q]];
id_and_value *order_perm_q = order_perm[q];
k = 1;
for (i=0;i<count[q];i++)
{
order_perm_q[i].id = perm_q[i];
order_perm_q[i].value = y[perm_q[i]];
}
qsort(order_perm_q, count[q], sizeof(id_and_value), compare_values);
int_y[order_perm_q[0].id] = 1;
for(i=1;i<count[q];i++)
{
if (order_perm_q[i-1].value<order_perm_q[i].value)
k++;
int_y[order_perm_q[i].id] = k;
}
nr_class[q] = k;
}
}
//
// Interface functions
//
model* train(const problem *prob, const parameter *param)
{
int l = prob->l;
int n = prob->n;
int w_size = prob->n;
model *model_ = Malloc(model,1);
model_->nr_feature=n;
model_->param = *param;
if(param->solver_type == SELECTION_TREE)
{
model_->w = Malloc(double, w_size);
model_->nr_class = 2;
int nr_subset;
int *start = NULL;
int *count = NULL;
int *perm = Malloc(int,l);
group_queries(prob, &nr_subset ,&start, &count, perm);
train_one(prob, param, &model_->w[0],0,0, nr_subset, perm, start, count);
free(start);
free(count);
free(perm);
}
void eval_list(double *label, double *target, int *query, int l, double *result_ret)
{
int q,i,j,k;
int nr_query;
int *start = NULL;
int *count = NULL;
int *perm = Malloc(int, l);
id_and_value *order_perm;
int true_query;
long long totalnc = 0, totalnd = 0;
long long nc = 0;
long long nd = 0;
double tmp;
double accuracy = 0;
int *l_plus;
int *int_y;
int same_y = 0;
double *ideal_dcg;
double *dcg;
double meanndcg = 0;
double ndcg;
selectiontree *T;
group_queries(query, l, &nr_query, &start, &count, perm);
true_query = nr_query;
for (q=0;q<nr_query;q++)
{
//We use selection trees to compute pairwise accuracy
nc = 0;
nd = 0;
l_plus = new int[count[q]];
int_y = new int[count[q]];
order_perm = new id_and_value[count[q]];
int *perm_q = &perm[start[q]];
for (i=0;i<count[q];i++)
{
order_perm[i].id = i;
order_perm[i].value = label[perm_q[i]];
}
qsort(order_perm, count[q], sizeof(id_and_value), compare_values);
int_y[order_perm[0].id] = 1;
same_y = 0;
k = 1;
for(i=1;i<count[q];i++)
{
if (order_perm[i-1].value < order_perm[i].value)
{
same_y = 0;
k++;
}
else
same_y++;
int_y[order_perm[i].id] = k;
nc += (i - same_y);
}
for (i=0;i<count[q];i++)
{
order_perm[i].id = i;
order_perm[i].value = target[perm_q[i]];
}
qsort(order_perm, count[q], sizeof(id_and_value), compare_values);
//total pairs
T = new selectiontree(k);
j = 0;
for (i=0;i<count[q];i++)
{
while (j<count[q] && ( order_perm[j].value < order_perm[i].value))
{
T->insert_node(int_y[order_perm[j].id], tmp);
j++;
}
T->larger(int_y[order_perm[i].id], &l_plus[order_perm[i].id], &tmp);
}
delete T;
for (i=0;i<count[q];i++)
nd += l_plus[i];
nc -= nd;
if (nc != 0 || nd != 0)
accuracy += double(nc)/double(nc+nd);
else
true_query--;
totalnc += nc;
totalnd += nd;
delete[] l_plus;
delete[] int_y;
delete[] order_perm;
}
result_ret[0] = (double)totalnc/(double)(totalnc+totalnd);
for (q=0;q<nr_query;q++)
{
//mean ndcg by the formulation of LETOR
ideal_dcg = new double[count[q]];
dcg = new double[count[q]];
ndcg = 0;
order_perm = new id_and_value[count[q]];
int *perm_q = &perm[start[q]];
for (i=0;i<count[q];i++)
{
order_perm[i].id = perm_q[i];
order_perm[i].value = label[perm_q[i]];
}
qsort(order_perm, count[q], sizeof(id_and_value), compare_values);
ideal_dcg[0] = pow(2.0,order_perm[count[q]-1].value) - 1;
for (i=1;i<count[q];i++)
ideal_dcg[i] = ideal_dcg[i-1] + (pow(2.0,order_perm[count[q]-1 - i].value) - 1) * log(2.0) / log(i+1.0);
for (i=0;i<count[q];i++)
{
order_perm[i].id = perm_q[i];
order_perm[i].value = target[perm_q[i]];
}
qsort(order_perm, count[q], sizeof(id_and_value), compare_values);
dcg[0] = pow(2.0, label[order_perm[count[q] - 1].id]) - 1;
for (i=1;i<count[q];i++)
dcg[i] = dcg[i-1] + (pow(2.0, label[order_perm[count[q] - 1 - i].id]) - 1) * log(2.0) / log(i + 1.0);
if (ideal_dcg[0]>0)
for (i=0;i<count[q];i++)
ndcg += dcg[i]/ideal_dcg[i];
else
ndcg = 0;
meanndcg += ndcg/count[q];
delete[] order_perm;
delete[] ideal_dcg;
delete[] dcg;
}
meanndcg /= nr_query;
result_ret[1] = meanndcg;
free(start);
free(count);
free(perm);
}
l2r_rank_fun::l2r_rank_fun(const rksvm_problem *prob, const rksvm_parameter *param,
Scheduler *scheduler, struct SolutionInfo *si)
{
this->si = si;
this->param = param;
si->rho = 0;
si->upper_bound_p = INF;
si->upper_bound_n = INF;
int l=prob->l;
this->prob = prob;
this->C = param->C;
this->thread_count = param->thread_count;//
this->current_rank = mpi_get_rank();//
this->global_l = prob->global_l;//
z = new double[l];
int i,j,k;
perm = new int[l];
group_queries(prob, &nr_subset ,&start, &count, perm);
pi = new id_and_value* [nr_subset];
#pragma omp parallel for default(shared) if(nr_subset > 50)
for (int i=0;i<nr_subset;i++)
{
pi[i] = new id_and_value[count[i]];
}
double *y=prob->y;
int_y = new int[prob->l];
nr_class = new int[nr_subset];
l_plus = new int[l];
l_minus = new int[l];
gamma_plus = new double[l];
gamma_minus = new double[l];
ATAQb = new double[l];
ATe = new double[l];
// the variable we have changed;
this->scheduler = scheduler;
this->local_l = scheduler->local_l;
this->start_ptr = scheduler->start_ptr;
//this->nr_recv = scheduler->nr_recv;
//this->nr_send = scheduler->nr_send;
gz = new double[global_l];
//gATAQb = new double[global_];
//gATe = new double[global_l];
Q = new double[l*global_l];
//here, it shows how to compute Q through TBB library.
nomad_fun(prob, param, scheduler, Q);
//testing Q
//char *file = "/home/jing/model/Q.txt";
//save_Q(file, prob, Q);
//mpi_exit(1);
#pragma omp parallel for default(shared) private(i,j,k)
for (i=0;i<nr_subset;i++)
{
k=1;
for (j=0;j<count[i];j++)
{
pi[i][j].id=perm[j+start[i]];
pi[i][j].value=y[perm[j+start[i]]];
}
qsort(pi[i], count[i], sizeof(id_and_value), compare_id_and_value);
int_y[pi[i][count[i]-1].id]=1;
for(j=count[i]-2;j>=0;j--)
{
if (pi[i][j].value>pi[i][j+1].value)
k++;
int_y[pi[i][j].id]=k;
}
nr_class[i]=k;
}
}