aboutsummaryrefslogtreecommitdiff
path: root/measures.R
blob: 99aa4df871e18a15420996a04113dc2f3fce31cd (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
stress <- function(Dx, Dy) {
  if (any(Dx != t(Dx)) || any(Dy != t(Dy))) {
    stop("Dx and Dy must be symmetric")
  }

  Dx <- as.matrix(Dx)
  Dy <- as.matrix(Dy)
  if (nrow(Dx) != nrow(Dy)) {
    stop("Dx and Dy must have the same number of elements")
  }

  n <- nrow(Dx)
  s <- vector("numeric", n)
  for (i in 1:n) {
    s[i] <- 0
    for (j in 1:n) {
      if (i == j) {
        next
      }

      s[i] = s[i] + (Dx[i, j] - Dy[i, j])^2 / Dx[i, j]
    }
    s[i] = s[i] / sum(D[i, ])
  }

  s
}

NP <- function(Dx, Dy, k = 9) {
  if (any(Dx != t(Dx)) || any(Dy != t(Dy))) {
    stop("Dx and Dy must be symmetric")
  }

  Dx <- as.matrix(Dx)
  Dy <- as.matrix(Dy)
  if (nrow(Dx) != nrow(Dy)) {
    stop("Dx and Dy must have the same number of elements")
  }

  n <- nrow(Dx)
  if (k >= n) {
    stop("k must be smaller than the number of elements")
  }

  preservation <- vector("numeric", n)
  for (i in 1:n) {
    nx <- order(Dx[i, ])[1 + 1:k]
    ny <- order(Dy[i, ])[1 + 1:k]
    diff <- setdiff(nx, ny)
    preservation[i] <- (k - length(diff)) / k
  }

  preservation
}

silhouette <- function(Dy, labels) {
  if (any(t(Dy) != Dy)) {
    stop("Dy must be symmetric")
  }

  Dy <- as.matrix(Dy)
  n <- nrow(Dy)
  cohesion <- vector("numeric", n)
  separation <- vector("numeric", n)

  for (i in 1:n) {
    label <- labels[i]
    separation[i] <- min(Dy[i, labels != label])
    cohesion[i] <- mean(Dy[i, labels[-i] == label])
  }

  silh <- (separation - cohesion) / max(separation, cohesion)
}

d2p <- function(D, perplexity = 9, tol = 1e-5, max.tries = 50) {
  if (any(D != t(D))) {
    stop("D must be symmetric")
  }

  D <- as.matrix(D)
  P <- matrix(data=0, nrow=nrow(D), ncol=ncol(D))
  n <- nrow(D)
  beta <- rep(1, n)
  logU <- log(perplexity)

  for (i in 1:n) {
    #denom <- sum(exp(-D[i, ] / sigmas))
    #P[i, ] <- exp(-D[i, ] / sigmas) / denom

    betaMin <- -Inf
    betaMax <-  Inf
    Di <- D[i, -i]

    tries <- 0
    repeat {
      Pi <- exp(-Di * beta[i])
      sumPi <- sum(Pi)
      H <- log(sumPi) + beta[i] * sum(Di * Pi) / sumPi
      Pi <- Pi / sumPi
      Hdiff <- H - logU

      if (abs(Hdiff) < tol || tries > max.tries) {
        break
      }

      if (Hdiff > 0) {
        betaMin <- beta[i]
        beta[i] <- if (is.finite(betaMax)) {
                     (beta[i] + betaMax) / 2
                   } else {
                     beta[i] * 2
                   }
      } else {
        betaMax <- beta[i]
        beta[i] <- if (is.finite(betaMin)) {
                     (beta[i] + betaMin) / 2
                   } else {
                     beta[i] / 2
                   }
      }

      tries <- tries + 1
    }

    P[i, -i] <- Pi
  }

  list(P = P, beta = beta) # sigmas = sqrt(1 / beta)
}

d2p.beta <- function(D, beta) {
  if (any(D != t(D))) {
    stop("D must be symmetric")
  }

  D <- as.matrix(D)
  n <- nrow(D)
  P <- matrix(data=0, nrow=nrow(D), ncol=ncol(D))
  for (i in 1:n) {
    P[i, -i] <- exp(-D[i, -i] * beta[i])
    P[i, -i] <- P[i, -i] / sum(P[i, -i])
  }

  P
}

klDivergence <- function(P, Q, eps = 1e-12) {
  if (nrow(P) != ncol(P) || nrow(Q) != ncol(Q)) {
    stop("P and Q must be square")
  }
  if (nrow(P) != nrow(Q)) {
    stop("P and Q must have the same number of elements")
  }

  P[P < eps] <- eps
  Q[Q < eps] <- eps
  n <- nrow(P)
  d <- vector("numeric", n)
  for (i in 1:n) {
    d[i] <- sum(P[i, -i] * log(P[i, -i] / Q[i, -i]))
  }

  d
}