func NewNaiveBayes(c *LabeledCorpus) *NaiveBayes {
labelCounts := make([]float64, c.LabelIndex.Len())
wordCounts := make([][]float64, c.LabelIndex.Len())
for i := range wordCounts {
wordCounts[i] = make([]float64, c.V)
for v := 0; v < c.V; v++ {
wordCounts[i][v]++ // smoothing
}
}
for d := 0; d < c.M; d++ {
label := c.LabelIndex.Id(c.Labels[d])
labelCounts[label]++
for n := 0; n < c.N[d]; n++ {
wordCounts[label][c.W[d][n]]++
}
}
util.Normalize(labelCounts)
util.Lcounts(labelCounts)
for _, counts := range wordCounts {
util.Normalize(counts)
util.Lcounts(counts)
}
return &NaiveBayes{c.LabelIndex, labelCounts, wordCounts}
}
func (i *ITM) Threshold(cutoff float64) {
for _, d := range rand.Perm(i.M) {
for n := 0; n < i.N[d]; n++ {
i.unsetZ(d, n)
counts := i.conditional(d, n)
util.Normalize(counts)
bestZ := -1
bestP := math.Inf(-1)
useBest := true
for z, count := range counts {
if count > bestP {
bestZ = z
bestP = count
}
if count < cutoff {
counts[z] = 0
} else {
useBest = false
}
}
if useBest {
i.setZ(d, n, bestZ)
} else {
i.setZ(d, n, util.SampleCounts(counts))
}
}
}
}
func FindAnchors(Q *matrix.DenseMatrix, k, r int) []int {
QRed := Q.Copy()
for _, row := range QRed.Arrays() {
util.Normalize(row)
}
QRed = QRed.Transpose()
QRed = RandomProjection(QRed, r)
QRed = QRed.Transpose()
return GramSchmidt(QRed, k)
}
