I have used a variety of tools for binary, multiclass and even incremental SVM problems, today I found something quite nice in binary case for libSVM, although potentially a source of confusion.

It is common in machine learning to apply a sigmoid function to normalise the boundaries of a problem, this can by empirically defining the upper and lower bound or through experimentation. Within libSVM they do this through experimentation, that is great to save you some time. The only thing to remember is it means through the use of random and cross validation with small sets of data, you are likely to get different results on each run.

So the function to consider is this:

// Cross-validation decision values for probability estimates
static void svm_binary_svc_probability(
const svm_problem *prob, const svm_parameter *param,
double Cp, double Cn, double& probA, double& probB)
{
int i;
int nr_fold = 5;
int *perm = Malloc(int,prob->l);
double *dec_values = Malloc(double,prob->l);
// random shuffle
for(i=0;i<prob->l;i++) perm[i]=i;
for(i=0;i<prob->l;i++)
{
int j = i+rand()%(prob->l-i);
swap(perm[i],perm[j]);
}
for(i=0;i<nr_fold;i++)
{
int begin = i*prob->l/nr_fold;
int end = (i+1)*prob->l/nr_fold;
int j,k;
struct svm_problem subprob;
subprob.l = prob->l-(end-begin);
subprob.x = Malloc(struct svm_node*,subprob.l);
subprob.y = Malloc(double,subprob.l);
k=0;
for(j=0;j<begin;j++)
{
subprob.x[k] = prob->x[perm[j]];
subprob.y[k] = prob->y[perm[j]];
++k;
}
for(j=end;j<prob->l;j++)
{
subprob.x[k] = prob->x[perm[j]];
subprob.y[k] = prob->y[perm[j]];
++k;
}
int p_count=0,n_count=0;
for(j=0;j<k;j++)
if(subprob.y[j]>0)
p_count++;
else
n_count++;
if(p_count==0 && n_count==0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 0;
else if(p_count > 0 && n_count == 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 1;
else if(p_count == 0 && n_count > 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = -1;
else
{
svm_parameter subparam = *param;
subparam.probability=0;
subparam.C=1.0;
subparam.nr_weight=2;
subparam.weight_label = Malloc(int,2);
subparam.weight = Malloc(double,2);
subparam.weight_label[0]=+1;
subparam.weight_label[1]=-1;
subparam.weight[0]=Cp;
subparam.weight[1]=Cn;
struct svm_model *submodel = svm_train(&subprob,&subparam);
for(j=begin;j<end;j++)
{
svm_predict_values(submodel,prob->x[perm[j]],&(dec_values[perm[j]]));
// ensure +1 -1 order; reason not using CV subroutine
dec_values[perm[j]] *= submodel->label[0];
}
svm_free_and_destroy_model(&submodel);
svm_destroy_param(&subparam);
}
free(subprob.x);
free(subprob.y);
}
sigmoid_train(prob->l,dec_values,prob->y,probA,probB);
free(dec_values);
free(perm);
}

So if you have few numbers of samples, that is the case in some circumstances then the cross validation is where you hit problems. Of course you can simply re-implement it yourself or you can add a few lines to stop cross validation if the number of samples is too few.

Not the most elegant of code, but for the moment it will do. I choose to completely seperate the two steps as opposed to multiple if’s

static void svm_binary_svc_probability(
const svm_problem *prob, const svm_parameter *param,
double Cp, double Cn, double& probA, double& probB)
{
int i;
int nr_fold = 5;
int *perm = Malloc(int,prob->l);
double *dec_values = Malloc(double,prob->l);
// random shuffle
for(i=0;i<prob->l;i++) perm[i]=i;
for(i=0;i<prob->l;i++)
{
int j = i+rand()%(prob->l-i);
swap(perm[i],perm[j]);
}
if (prob->l < (5*nr_fold)){
int begin = 0;
int end = prob->l;
int j,k;
struct svm_problem subprob;
subprob.l = prob->l;
subprob.x = Malloc(struct svm_node*,subprob.l);
subprob.y = Malloc(double,subprob.l);
k=0;
for(j=0;j<prob->l;j++)
{
subprob.x[k] = prob->x[perm[j]];
subprob.y[k] = prob->y[perm[j]];
++k;
}
int p_count=0,n_count=0;
for(j=0;j<k;j++)
if(prob->y[j]>0)
p_count++;
else
n_count++;
if(p_count==0 && n_count==0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 0;
else if(p_count > 0 && n_count == 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 1;
else if(p_count == 0 && n_count > 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = -1;
else
{
svm_parameter subparam = *param;
subparam.probability=0;
subparam.C=1.0;
subparam.nr_weight=2;
subparam.weight_label = Malloc(int,2);
subparam.weight = Malloc(double,2);
subparam.weight_label[0]=+1;
subparam.weight_label[1]=-1;
subparam.weight[0]=Cp;
subparam.weight[1]=Cn;
struct svm_model *submodel = svm_train(&subprob,&subparam);
for(j=begin;j<end;j++)
{
svm_predict_values(submodel,prob->x[perm[j]],&(dec_values[perm[j]]));
// ensure +1 -1 order; reason not using CV subroutine
dec_values[perm[j]] *= submodel->label[0];
}
svm_free_and_destroy_model(&submodel);
svm_destroy_param(&subparam);
}
free(subprob.x);
free(subprob.y);
}else{
for(i=0;i<nr_fold;i++)
{
int begin = i*prob->l/nr_fold;
int end = (i+1)*prob->l/nr_fold;
if (nr_fold == 1){
begin = 0 ;
end = prob->l;
}
int j,k;
struct svm_problem subprob;
subprob.l = prob->l-(end-begin);
subprob.x = Malloc(struct svm_node*,subprob.l);
subprob.y = Malloc(double,subprob.l);
k=0;
for(j=0;j<begin;j++)
{
subprob.x[k] = prob->x[perm[j]];
subprob.y[k] = prob->y[perm[j]];
++k;
}
for(j=end;j<prob->l;j++)
{
subprob.x[k] = prob->x[perm[j]];
subprob.y[k] = prob->y[perm[j]];
++k;
}
int p_count=0,n_count=0;
for(j=0;j<k;j++)
if(subprob.y[j]>0)
p_count++;
else
n_count++;
if(p_count==0 && n_count==0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 0;
else if(p_count > 0 && n_count == 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = 1;
else if(p_count == 0 && n_count > 0)
for(j=begin;j<end;j++)
dec_values[perm[j]] = -1;
else
{
svm_parameter subparam = *param;
subparam.probability=0;
subparam.C=1.0;
subparam.nr_weight=2;
subparam.weight_label = Malloc(int,2);
subparam.weight = Malloc(double,2);
subparam.weight_label[0]=+1;
subparam.weight_label[1]=-1;
subparam.weight[0]=Cp;
subparam.weight[1]=Cn;
struct svm_model *submodel = svm_train(&subprob,&subparam);
for(j=begin;j<end;j++)
{
svm_predict_values(submodel,prob->x[perm[j]],&(dec_values[perm[j]]));
// ensure +1 -1 order; reason not using CV subroutine
dec_values[perm[j]] *= submodel->label[0];
}
svm_free_and_destroy_model(&submodel);
svm_destroy_param(&subparam);
}
free(subprob.x);
free(subprob.y);
}
}
sigmoid_train(prob->l,dec_values,prob->y,probA,probB);
free(dec_values);
free(perm);
}

As with a lot of my code based posts, this is more for my memory than anything, but hopefully may help people unlock the secrets of libSVM.