epclust/tests/testthat/test.clustering.R

   1 context("clustering")
   2
   3 #shorthand: map 1->1, 2->2, 3->3, 4->1, ..., 149->2, 150->3, ... (is base==3)
   4 I = function(i, base)
   5     (i-1) %% base + 1
   6
   7 test_that("computeClusters1 behave as expected",
   8 {
   9     require("MASS", quietly=TRUE)
  10     if (!require("clue", quietly=TRUE))
  11         skip("'clue' package not available")
  12
  13     # 3 gaussian clusters, 300 items; and then 7 gaussian clusters, 490 items
  14     n = 300
  15     d = 5
  16     K = 3
  17     for (ndK in list( c(300,5,3), c(490,10,7) ))
  18     {
  19         n = ndK[1] ; d = ndK[2] ; K = ndK[3]
  20         cs = n/K #cluster size
  21         Id = diag(d)
  22         coefs = do.call(rbind,
  23             lapply(1:K, function(i) MASS::mvrnorm(cs, c(rep(0,(i-1)),5,rep(0,d-i)), Id)))
  24         indices_medoids = computeClusters1(coefs, K)
  25         # Get coefs assignments (to medoids)
  26         assignment = sapply(seq_len(n), function(i)
  27             which.min( rowSums( sweep(coefs[indices_medoids,],2,coefs[i,],'-')^2 ) ) )
  28         for (i in 1:K)
  29             expect_equal(sum(assignment==i), cs, tolerance=5)
  30
  31         costs_matrix = matrix(nrow=K,ncol=K)
  32         for (i in 1:K)
  33         {
  34             for (j in 1:K)
  35             {
  36                 # assign i (in result) to j (order 1,2,3)
  37                 costs_matrix[i,j] = abs( mean(assignment[((i-1)*cs+1):(i*cs)]) - j )
  38             }
  39         }
  40         permutation = as.integer( clue::solve_LSAP(costs_matrix) )
  41         for (i in 1:K)
  42         {
  43             expect_equal(
  44                 mean(assignment[((i-1)*cs+1):(i*cs)]), permutation[i], tolerance=0.05)
  45         }
  46     }
  47 })
  48
  49 test_that("computeSynchrones behave as expected",
  50 {
  51     n = 300
  52     x = seq(0,9.5,0.1)
  53     L = length(x) #96 1/4h
  54     K = 3
  55     s1 = cos(x)
  56     s2 = sin(x)
  57     s3 = c( s1[1:(L%/%2)] , s2[(L%/%2+1):L] )
  58     #sum((s1-s2)^2) == 96
  59     #sum((s1-s3)^2) == 58
  60     #sum((s2-s3)^2) == 38
  61     s = list(s1, s2, s3)
  62     series = matrix(nrow=n, ncol=L)
  63     for (i in seq_len(n))
  64         series[i,] = s[[I(i,K)]] + rnorm(L,sd=0.01)
  65     getRefSeries = function(indices) {
  66         indices = indices[indices < n]
  67         if (length(indices)>0) series[indices,] else NULL
  68     }
  69     synchrones = computeSynchrones(rbind(s1,s2,s3), getRefSeries, 100)
  70
  71     expect_equal(dim(synchrones), c(K,L))
  72     for (i in 1:K)
  73         expect_equal(synchrones[i,], s[[i]], tolerance=0.01)
  74 })
  75
  76 computeDistortion = function(series, medoids)
  77 {
  78     n = nrow(series) ; L = ncol(series)
  79     distortion = 0.
  80     for (i in seq_len(n))
  81         distortion = distortion + min( rowSums( sweep(medoids,2,series[i,],'-')^2 ) / L )
  82     distortion / n
  83 }
  84
  85 test_that("computeClusters2 behave as expected",
  86 {
  87     n = 900
  88     x = seq(0,9.5,0.1)
  89     L = length(x) #96 1/4h
  90     K1 = 60
  91     K2 = 3
  92     #for (i in 1:60) {plot(x^(1+i/30)*cos(x+i),type="l",col=i,ylim=c(-50,50)); par(new=TRUE)}
  93     s = lapply( seq_len(K1), function(i) x^(1+i/30)*cos(x+i) )
  94     series = matrix(nrow=n, ncol=L)
  95     for (i in seq_len(n))
  96         series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01)
  97     getRefSeries = function(indices) {
  98         indices = indices[indices < n]
  99         if (length(indices)>0) series[indices,] else NULL
 100     }
 101     # Artificially simulate 60 medoids - perfect situation, all equal to one of the refs
 102     medoids_K1 = do.call(rbind, lapply( 1:K1, function(i) s[[I(i,K1)]] ) )
 103     medoids_K2 = computeClusters2(medoids_K1, K2, getRefSeries, 75)
 104
 105     expect_equal(dim(medoids_K2), c(K2,L))
 106     # Not easy to evaluate result: at least we expect it to be better than random selection of
 107     # medoids within 1...K1 (among references)
 108
 109     distorGood = computeDistortion(series, medoids_K2)
 110     for (i in 1:3)
 111         expect_lte( distorGood, computeDistortion(series,medoids_K1[sample(1:K1, K2),]) )
 112 })
 113
 114 test_that("clusteringTask + computeClusters2 behave as expected",
 115 {
 116     n = 900
 117     x = seq(0,9.5,0.1)
 118     L = length(x) #96 1/4h
 119     K1 = 60
 120     K2 = 3
 121     s = lapply( seq_len(K1), function(i) x^(1+i/30)*cos(x+i) )
 122     series = matrix(nrow=n, ncol=L)
 123     for (i in seq_len(n))
 124         series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01)
 125     getSeries = function(indices) {
 126         indices = indices[indices <= n]
 127         if (length(indices)>0) series[indices,] else NULL
 128     }
 129     wf = "haar"
 130     ctype = "absolute"
 131     getContribs = function(indices) curvesToContribs(series[indices,],wf,ctype)
 132     medoids_K1 = getSeries( clusteringTask(1:n, getContribs, K1, 75, 4) )
 133     medoids_K2 = computeClusters2(medoids_K1, K2, getSeries, 120)
 134
 135     expect_equal(dim(medoids_K1), c(K1,L))
 136     expect_equal(dim(medoids_K2), c(K2,L))
 137     # Not easy to evaluate result: at least we expect it to be better than random selection of
 138     # medoids within 1...K1 (among references)
 139     distorGood = computeDistortion(series, medoids_K2)
 140     for (i in 1:3)
 141         expect_lte( distorGood, computeDistortion(series,medoids_K1[sample(1:K1, K2),]) )
 142 })