/*
populate will generate new synthetic sample using nearest neighbors.
*/
func (smote *Runtime) populate(instance *tabula.Row, neighbors knn.Neighbors) {
lenAttr := len(*instance)
for x := 0; x < smote.NSynthetic; x++ {
// choose one of the K nearest neighbors
n := rand.Intn(neighbors.Len())
sample := neighbors.Row(n)
newSynt := make(tabula.Row, lenAttr)
// Compute new synthetic attributes.
for attr, sr := range *sample {
if attr == smote.ClassIndex {
continue
}
ir := (*instance)[attr]
iv := ir.Float()
sv := sr.Float()
dif := sv - iv
gap := rand.Float64()
newAttr := iv + (gap * dif)
record := &tabula.Record{}
record.SetFloat(newAttr)
newSynt[attr] = record
}
newSynt[smote.ClassIndex] = (*instance)[smote.ClassIndex]
smote.Synthetics.PushRow(&newSynt)
}
}
//
// createSynthetic will create synthetics row from original row `p` and their
// `neighbors`.
//
func (in *Runtime) createSynthetic(p *tabula.Row, neighbors knn.Neighbors) (
synthetic *tabula.Row,
) {
// choose one of the K nearest neighbors
randIdx := rand.Intn(neighbors.Len())
n := neighbors.Row(randIdx)
// Check if synthetic sample can be created from p and n.
canit, slp, sln := in.canCreate(p, n)
if !canit {
if DEBUG >= 2 {
fmt.Println("[lnsmote] can not create synthetic")
}
if slp.Len() <= 0 {
in.outliers.PushBack(p)
}
// we can not create from p and synthetic.
return nil
}
synthetic = p.Clone()
for x, srec := range *synthetic {
// Skip class attribute.
if x == in.ClassIndex {
continue
}
delta := in.randomGap(p, n, slp.Len(), sln.Len())
pv := (*p)[x].Float()
diff := (*n)[x].Float() - pv
srec.SetFloat(pv + delta*diff)
}
return
}
