public static void Quadratic(
int imin, int imax, int q,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
// Assumption: each cluster must have at least one point.
for (int i = imin; i <= imax; ++i)
{
S[q][i] = S[q - 1][i - 1];
J[q][i] = i;
int jmin = Math.Max(q, (int)J[q - 1][i]);
for (int j = i - 1; j >= jmin; --j)
{
number Sj = S[q - 1][j - 1] + WithinCluster.Dissimilarity(criterion, j, i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// ssq(j, i, sum_x, sum_x_sq, sum_w)
if (Sj < S[q][i])
{
S[q][i] = Sj;
J[q][i] = j;
}
}
}
}
public static void SMAWK(
int imin, int imax, int q,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
int[] js = new int[imax - q + 1];
int abs = (q);
for (int iter = 0; iter < js.Length; iter++)
{
js[iter] = abs++;
}
SMAWK(imin, imax, 1, q, js, S, J, sum_x, sum_x_sq, sum_w, sum_w_sq, criterion);
}
private static void FindMinFromCandidates(
int imin, int imax, int istep, int q,
int[] js,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
int rmin_prev = 0;
for (int i = (imin); i <= imax; i += istep)
{
int rmin = rmin_prev;
// Initialization of S[q][i] and J[q][i]
S[q][i] = S[q - 1][js[rmin] - 1] + WithinCluster.Dissimilarity(criterion, js[rmin], i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// ssq(js[rmin], i, sum_x, sum_x_sq, sum_w);
J[q][i] = js[rmin];
for (int r = (rmin + 1); r < js.Length; ++r)
{
int j_abs = js[r];
if (j_abs < J[q - 1][i])
{
continue;
}
if (j_abs > i)
{
break;
}
number Sj = (S[q - 1][j_abs - 1] + WithinCluster.Dissimilarity(criterion, j_abs, i, sum_x, sum_x_sq, sum_w, sum_w_sq));
// ssq(j_abs, i, sum_x, sum_x_sq, sum_w));
if (Sj <= S[q][i])
{
S[q][i] = Sj;
J[q][i] = js[r];
rmin_prev = r;
}
}
}
}
public static number Dissimilarity(DissimilarityType disType, int j, int i,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq = null)
{
number d = 0;
switch (disType)
{
case DissimilarityType.L1:
d = SABS(j, i, sum_x, sum_w);
break;
case DissimilarityType.L2:
d = SSQ(j, i, sum_x, sum_x_sq, sum_w);
break;
case DissimilarityType.L2Y:
d = SSQ(j, i, sum_w, sum_w_sq);
break;
}
return(d);
}
public static void FillMatrix(number[] x, number[] w, number[][] S, int[][] J, Method method, DissimilarityType criterion)
{
/*
* x: One dimension vector to be clustered, must be sorted (in any order).
* S: K x N matrix. S[q][i] is the sum of squares of the distance from
* each x[i] to its cluster mean when there are exactly x[i] is the
* last point in cluster q
* J: K x N backtrack matrix
*
* NOTE: All vector indices in this program start at position 0
*/
int K = S.Length;
int N = S[0].Length;
number[] sum_x = new number[N];
number[] sum_x_sq = new number[N];
number[] sum_w = null;
number[] sum_w_sq = null;
int[] jseq = new int[N];
number shift = x[N / 2]; // median. used to shift the values of x to
// improve numerical stability
if (w == null || w.Length == 0)
{
// equally weighted
sum_x[0] = x[0] - shift;
sum_x_sq[0] = (x[0] - shift) * (x[0] - shift);
}
else
{ // unequally weighted
sum_x[0] = w[0] * (x[0] - shift);
sum_x_sq[0] = w[0] * (x[0] - shift) * (x[0] - shift);
sum_w = new number[N];
sum_w_sq = new number[N];
sum_w[0] = w[0];
sum_w_sq[0] = w[0] * w[0];
}
S[0][0] = 0;
J[0][0] = 0;
for (int i = 1; i < N; ++i)
{
if (w == null || w.Length == 0)
{ // equally weighted
sum_x[i] = sum_x[i - 1] + x[i] - shift;
sum_x_sq[i] = sum_x_sq[i - 1] + (x[i] - shift) * (x[i] - shift);
}
else
{ // unequally weighted
sum_x[i] = sum_x[i - 1] + w[i] * (x[i] - shift);
sum_x_sq[i] = sum_x_sq[i - 1] + w[i] * (x[i] - shift) * (x[i] - shift);
sum_w[i] = sum_w[i - 1] + w[i];
sum_w_sq[i] = sum_w_sq[i - 1] + w[i] * w[i];
}
// Initialize for q = 0
S[0][i] = WithinCluster.Dissimilarity(criterion, 0, i, sum_x, sum_x_sq, sum_w, sum_w_sq); // ssq(0, i, sum_x, sum_x_sq, sum_w);
J[0][i] = 0;
}
/*
#if DEBUG
* for (int i = 0; i < x.Length; ++i)
* {
* Console.Write(x[i] + ",");
* }
* Console.WriteLine();
#endif
*/
for (int q = 1; q < K; ++q)
{
int imin;
if (q < K - 1)
{
imin = Math.Max(1, q);
}
else
{
// No need to compute S[K-1][0] ... S[K-1][N-2]
imin = N - 1;
}
/*
# ifdef DEBUG
# // std::cout << std::endl << "q=" << q << ":" << std::endl;
#endif
*/
// fill_row_k_linear_recursive(imin, N-1, 1, q, jseq, S, J, sum_x, sum_x_sq);
// fill_row_k_linear(imin, N-1, q, S, J, sum_x, sum_x_sq);
if (method == Method.Linear)
{
Fill.SMAWK(imin, N - 1, q, S, J, sum_x, sum_x_sq, sum_w, sum_w_sq, criterion);
}
else if (method == Method.LogLinear)
{
Fill.LogLinear(imin, N - 1, q, q, N - 1, S, J, sum_x, sum_x_sq, sum_w, sum_w_sq, criterion);
}
else if (method == Method.Quadratic)
{
Fill.Quadratic(imin, N - 1, q, S, J, sum_x, sum_x_sq, sum_w, sum_w_sq, criterion);
}
else
{
throw new Exception("ERROR: unknown method " + method + "!");
}
/*
#if DEBUG
*
* fill_row_q_log_linear(imin, N - 1, q, q, N - 1, SS, JJ, sum_x, sum_x_sq, sum_w, sum_w_sq, criterion);
*
* for (int i = imin; i < N; ++i)
* {
* if (S[q][i] != SS[q][i] || J[q][i] != JJ[q][i])
* {
* std::cout << "ERROR: q=" << q << ", i=" << i << std::endl;
* std::cout << "\tS=" << S[q][i] << "\tJ=" << J[q][i] << std::endl;
* std::cout << "Truth\tSS=" << SS[q][i] << "\tJJ=" << JJ[q][i];
* std::cout << std::endl;
* assert(false);
*
* }
* else
* {
*
* std::cout << "OK: q=" << q << ", i=" << i << std::endl;
* std::cout << "\tS=" << S[q][i] << "\tJ=" << J[q][i] << std::endl;
* std::cout << "Truth\tSS=" << SS[q][i] << "\tJJ=" << JJ[q][i];
* std::cout << std::endl;
*
* }
*
* }
#endif
*/
}
/*
# ifdef DEBUG
# std::cout << "Linear & log-linear code returned identical dp index matrix."
# << std::endl;
#endif
*/
}
private static void KMeans(number[] x, number[] y, int Kmin, int Kmax, out int[] clusters, out number[] centers, out number[] withinss, out number[] size, out double[] BIC, Method method, DissimilarityType criterion)
{
// Input:
// x -- an array of double precision numbers, not necessarily sorted
// Kmin -- the minimum number of clusters expected
// Kmax -- the maximum number of clusters expected
// NOTE: All vectors in this program is considered starting at position 0.
int N = x.Length;
clusters = new int[N];
BIC = new double[Kmax - Kmin];
int[] order = new int[N];
for (int i = 0; i < order.Length; ++i)
{
order[i] = i;
}
bool is_sorted = true;
for (int i = 0; i < N - 1; ++i)
{
if (x[i] > x[i + 1])
{
is_sorted = false;
break;
}
}
number[] x_sorted = null;
number[] y_sorted = null;
bool is_equally_weighted = true;
if (!is_sorted)
{
x_sorted = new number[x.Length];
Array.Copy(x, x_sorted, x.Length);
Array.Sort(x_sorted, order);
for (int i = 0; i < x_sorted.Length; i++)
{
x_sorted[i] = x[order[i]];
}
}
else
{
x_sorted = x;
}
// check to see if unequal weight is provided
if (y != null)
{
is_equally_weighted = true;
for (int i = 1; i < N; ++i)
{
if (y[i] != y[i - 1])
{
is_equally_weighted = false;
break;
}
}
}
if (!is_equally_weighted)
{
y_sorted = new number[N];
for (int i = 0; i < N; ++i)
{
y_sorted[i] = y[order[i]];
}
}
else
{
y = null;
}
int nUnique = 1;
if (N == 0)
{
nUnique = 0;
}
if (N > 1)
{
for (int i = 1; i < N; i++)
{
if (x_sorted[i - 1] != x_sorted[i])
{
nUnique++;
}
}
}
Kmax = nUnique < Kmax ? nUnique : Kmax;
if (nUnique > 1)
{ // The case when not all elements are equal.
number[][] S = new number[Kmax][];
for (int i = 0; i < Kmax; i++)
{
S[i] = new number[N];
}
int[][] J = new int[Kmax][];
for (int i = 0; i < Kmax; i++)
{
J[i] = new int[N];
}
int Kopt;
DynamicProgramming.FillMatrix(x_sorted, y_sorted, S, J, method, criterion);
// Fill in dynamic programming matrix
if (is_equally_weighted)
{
Kopt = NonWeighted.SelectLevels(x_sorted, J, Kmin, Kmax, BIC);
}
else
{
switch (criterion)
{
case DissimilarityType.L2Y:
Kopt = NonWeighted.SelectLevels(y_sorted, J, Kmin, Kmax, BIC);
break;
default:
Kopt = Weighted.SelectLevels(x_sorted, y_sorted, J, Kmin, Kmax, BIC);
break;
}
}
centers = new number[Kopt];
withinss = new number[Kopt];
size = new number[Kopt];
if (Kopt < Kmax)
{ // Reform the dynamic programming matrix S and J
Array.Resize(ref J, Kopt);
}
int[] cluster_sorted = new int[N];
// Backtrack to find the clusters beginning and ending indices
if (is_equally_weighted && criterion == DissimilarityType.L1)
{
DynamicProgramming.BacktrackL1(x_sorted, J, cluster_sorted, centers, withinss, size);
}
else if (is_equally_weighted && criterion == DissimilarityType.L2)
{
DynamicProgramming.Backtrack(x_sorted, J, cluster_sorted, centers, withinss, size);
}
else if (criterion == DissimilarityType.L2Y)
{
DynamicProgramming.BacktrackL2Y(x_sorted, y_sorted, J, cluster_sorted, centers, withinss, size);
}
else
{
DynamicProgramming.BacktrackWeighted(x_sorted, y_sorted, J, cluster_sorted, centers, withinss, size);
}
/*#if DEBUG
* std::cout << "backtrack done." << std::endl;
#endif*/
for (int i = 0; i < N; ++i)
{
// Obtain clustering on data in the original order
clusters[order[i]] = cluster_sorted[i];
}
}
else
{
// A single cluster that contains all elements
for (int i = 0; i < N; ++i)
{
clusters[i] = 0;
}
centers = new number[1];
withinss = new number[1];
size = new number[1];
centers[0] = x[0];
withinss[0] = 0;
size[0] = N * (is_equally_weighted ? 1 : y[0]);
}
}
public static void LogLinear(
int imin, int imax, int q,
int jmin, int jmax,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
if (imin > imax)
{
return;
}
int N = S[0].Length;
int i = (imin + imax) / 2;
#if DEBUG
// std::cout << " i=" << i << ": ";
#endif
// Initialization of S[q][i]:
S[q][i] = S[q - 1][i - 1];
J[q][i] = i;
int jlow = q; // the lower end for j
if (imin > q)
{
// jlow = std::max(jlow, (int)J[q][imin-1]);
jlow = Math.Max(jlow, jmin);
}
jlow = Math.Max(jlow, J[q - 1][i]);
int jhigh = i - 1; // the upper end for j
if (imax < N - 1)
{
// jhigh = std::min(jhigh, (int)J[q][imax+1]);
jhigh = Math.Min(jhigh, jmax);
}
#if DEBUG
// std::cout << " j-=" << jlow << ", j+=" << jhigh << ": ";
#endif
for (int j = jhigh; j >= jlow; --j)
{
// compute s(j,i)
number sji = WithinCluster.SSQ(j, i, sum_x, sum_x_sq, sum_w);
// MS May 11, 2016 Added:
if (sji + S[q - 1][jlow - 1] >= S[q][i])
{
break;
}
// Examine the lower bound of the cluster border
// compute s(jlow, i)
number sjlowi = WithinCluster.Dissimilarity(criterion, jlow, i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// ssq(jlow, i, sum_x, sum_x_sq, sum_w);
number SSQ_jlow = sjlowi + S[q - 1][jlow - 1];
if (SSQ_jlow < S[q][i])
{
// shrink the lower bound
S[q][i] = SSQ_jlow;
J[q][i] = jlow;
}
jlow++;
number SSQ_j = sji + S[q - 1][j - 1];
if (SSQ_j < S[q][i])
{
S[q][i] = SSQ_j;
J[q][i] = j;
}
}
#if DEBUG
//std::cout << // " q=" << q << ": " <<
// "\t" << S[q][i] << "\t" << J[q][i];
//std::cout << std::endl;
#endif
jmin = (imin > q) ? (int)J[q][imin - 1] : q;
jmax = (int)J[q][i];
LogLinear(imin, i - 1, q, jmin, jmax,
S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion);
jmin = (int)J[q][i];
jmax = (imax < N - 1) ? (int)J[q][imax + 1] : imax;
LogLinear(i + 1, imax, q, jmin, jmax,
S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion);
}
private static void SMAWK(
int imin, int imax, int istep, int q,
int[] js,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
#if DEBUG_REDUCE
std::cout << "i:" << '[' << imin << ',' << imax << ']' << '+' << istep
<< std::endl;
#endif
if (imax - imin <= 0 * istep)
{ // base case only one element left
FindMinFromCandidates(
imin, imax, istep, q, js, S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion
);
}
else
{
// REDUCE
#if DEBUG_REDUCE
std::cout << "js:";
for (size_t l = 0; l < js.size(); ++l)
{
std::cout << js[l] << ",";
}
std::cout << std::endl;
std::cout << std::endl;
#endif
int[] js_odd;
ReduceInPlace(imin, imax, istep, q, js, out js_odd,
S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion);
int istepx2 = (istep << 1);
int imin_odd = (imin + istep);
int imax_odd = (imin_odd + (imax - imin_odd) / istepx2 * istepx2);
// Recursion on odd rows (0-based):
SMAWK(imin_odd, imax_odd, istepx2,
q, js_odd, S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion);
#if DEBUG_REDUCE
std::cout << "js_odd (reduced):";
for (size_t l = 0; l < js_odd.size(); ++l)
{
std::cout << js_odd[l] << ",";
}
std::cout << std::endl << std::endl;
std::cout << "even pos:";
for (int i = imin; i < imax; i += istepx2)
{
std::cout << i << ",";
}
std::cout << std::endl << std::endl;
#endif
FillEvenPositions(imin, imax, istep, q, js,
S, J, sum_x, sum_x_sq, sum_w,
sum_w_sq, criterion);
}
}
private static void FillEvenPositions(
int imin, int imax, int istep, int q,
int[] js,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
// Derive j for even rows (0-based)
int n = js.Length;
int istepx2 = (istep << 1);
int jl = js[0];
for (int i = imin, r = 0; i <= imax; i += istepx2)
{
// auto jmin = (i == imin) ? js[0] : J[q][i - istep];
while (js[r] < jl)
{
// Increase r until it points to a value of at least jmin
r++;
}
// Initialize S[q][i] and J[q][i]
S[q][i] = S[q - 1][js[r] - 1] +
WithinCluster.Dissimilarity(criterion, js[r], i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// ssq(js[r], i, sum_x, sum_x_sq, sum_w);
J[q][i] = js[r]; // rmin
// Look for minimum S upto jmax within js
int jh = (i + istep <= imax)
? J[q][i + istep] : js[n - 1];
int jmax = Math.Min(jh, i);
number sjimin = WithinCluster.Dissimilarity(criterion, jmax, i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// ssq(jmax, i, sum_x, sum_x_sq, sum_w)
for (++r; r < n && js[r] <= jmax; r++)
{
int jabs = js[r];
if (jabs > i)
{
break;
}
if (jabs < J[q - 1][i])
{
continue;
}
number s = WithinCluster.Dissimilarity(criterion, jabs, i, sum_x, sum_x_sq, sum_w, sum_w_sq);
// (ssq(jabs, i, sum_x, sum_x_sq, sum_w));
number Sj = (S[q - 1][jabs - 1] + s);
if (Sj <= S[q][i])
{
S[q][i] = Sj;
J[q][i] = js[r];
}
else if (S[q - 1][jabs - 1] + sjimin > S[q][i])
{
break;
}
/*else if(S[q-1][js[rmin]-1] + s > S[q][i]) {
* break;
* }*/
}
r--;
jl = jh;
}
}
//SMAWK
private static void ReduceInPlace(
int imin, int imax, int istep, int q,
int[] js, out int[] js_red,
number[][] S, int[][] J,
number[] sum_x, number[] sum_x_sq, number[] sum_w, number[] sum_w_sq,
DissimilarityType criterion)
{
int N = (imax - imin) / istep + 1;
js_red = js;
if (N >= js.Length)
{
return;
}
// Two positions to move candidate j's back and forth
int left = -1; // points to last favorable position / column
int right = 0; // points to current position / column
int m = js_red.Length;
while (m > N)
{ // js_reduced has more than N positions / columns
int p = (left + 1);
int i = (imin + p * istep);
int j = (js_red[right]);
number Sl = (S[q - 1][j - 1] + WithinCluster.Dissimilarity(criterion, j, i, sum_x, sum_x_sq, sum_w, sum_w_sq));
// ssq(j, i, sum_x, sum_x_sq, sum_w));
int jplus1 = (js_red[right + 1]);
number Slplus1 = (S[q - 1][jplus1 - 1] + WithinCluster.Dissimilarity(criterion, jplus1, i, sum_x, sum_x_sq, sum_w, sum_w_sq));
// ssq(jplus1, i, sum_x, sum_x_sq, sum_w));
if (Sl < Slplus1 && p < N - 1)
{
js_red[++left] = j; // i += istep;
right++; // move on to next position / column p+1
}
else if (Sl < Slplus1 && p == N - 1)
{
js_red[++right] = j; // delete position / column p+1
m--;
}
else
{ // (Sl >= Slplus1)
if (p > 0)
{ // i > imin
// delete position / column p and
// move back to previous position / column p-1:
js_red[right] = js_red[left--];
// p --; // i -= istep;
}
else
{
right++; // delete position / column 0
}
m--;
}
}
for (int r = (left + 1); r < m; ++r)
{
js_red[r] = js_red[right++];
}
Array.Resize(ref js_red, m);
return;
}
