context("clustering") #shorthand: map 1->1, 2->2, 3->3, 4->1, ..., 149->2, 150->3, ... (is base==3) I = function(i, base) (i-1) %% base + 1 test_that("computeClusters1 behave as expected", { require("MASS", quietly=TRUE) require("clue", quietly=TRUE) # 3 gaussian clusters, 300 items; and then 7 gaussian clusters, 490 items n = 300 d = 5 K = 3 for (ndK in list( c(300,5,3), c(490,10,7) )) { n = ndK[1] ; d = ndK[2] ; K = ndK[3] cs = n/K #cluster size Id = diag(d) coefs = do.call(rbind, lapply(1:K, function(i) MASS::mvrnorm(cs, c(rep(0,(i-1)),5,rep(0,d-i)), Id))) indices_medoids = computeClusters1(coefs, K) # Get coefs assignments (to medoids) assignment = sapply(seq_len(n), function(i) which.min( rowSums( sweep(coefs[indices_medoids,],2,coefs[i,],'-')^2 ) ) ) for (i in 1:K) expect_equal(sum(assignment==i), cs, tolerance=5) costs_matrix = matrix(nrow=K,ncol=K) for (i in 1:K) { for (j in 1:K) { # assign i (in result) to j (order 1,2,3) costs_matrix[i,j] = abs( mean(assignment[((i-1)*cs+1):(i*cs)]) - j ) } } permutation = as.integer( clue::solve_LSAP(costs_matrix) ) for (i in 1:K) { expect_equal( mean(assignment[((i-1)*cs+1):(i*cs)]), permutation[i], tolerance=0.05) } } }) test_that("computeSynchrones behave as expected", { n = 300 x = seq(0,9.5,0.1) L = length(x) #96 1/4h K = 3 s1 = cos(x) s2 = sin(x) s3 = c( s1[1:(L%/%2)] , s2[(L%/%2+1):L] ) #sum((s1-s2)^2) == 96 #sum((s1-s3)^2) == 58 #sum((s2-s3)^2) == 38 s = list(s1, s2, s3) series = matrix(nrow=n, ncol=L) for (i in seq_len(n)) series[i,] = s[[I(i,K)]] + rnorm(L,sd=0.01) getRefSeries = function(indices) { indices = indices[indices < n] if (length(indices)>0) series[indices,] else NULL } synchrones = computeSynchrones(rbind(s1,s2,s3), getRefSeries, 100) expect_equal(dim(synchrones), c(K,L)) for (i in 1:K) expect_equal(synchrones[i,], s[[i]], tolerance=0.01) }) computeDistortion = function(series, medoids) { n = nrow(series) ; L = ncol(series) distortion = 0. for (i in seq_len(n)) distortion = distortion + min( rowSums( sweep(medoids,2,series[i,],'-')^2 ) / L ) distortion / n } test_that("computeClusters2 behave as expected", { n = 900 x = seq(0,9.5,0.1) L = length(x) #96 1/4h K1 = 60 K2 = 3 #for (i in 1:60) {plot(x^(1+i/30)*cos(x+i),type="l",col=i,ylim=c(-50,50)); par(new=TRUE)} s = lapply( seq_len(K1), function(i) x^(1+i/30)*cos(x+i) ) series = matrix(nrow=n, ncol=L) for (i in seq_len(n)) series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01) getRefSeries = function(indices) { indices = indices[indices < n] if (length(indices)>0) series[indices,] else NULL } # Artificially simulate 60 medoids - perfect situation, all equal to one of the refs medoids_K1 = do.call(rbind, lapply( 1:K1, function(i) s[[I(i,K1)]] ) ) medoids_K2 = computeClusters2(medoids_K1, K2, getRefSeries, 75) expect_equal(dim(medoids_K2), c(K2,L)) # Not easy to evaluate result: at least we expect it to be better than random selection of # medoids within 1...K1 (among references) distorGood = computeDistortion(series, medoids_K2) for (i in 1:3) expect_lte( distorGood, computeDistortion(series,medoids_K1[sample(1:K1, K2),]) ) }) test_that("clusteringTask + computeClusters2 behave as expected", { n = 900 x = seq(0,9.5,0.1) L = length(x) #96 1/4h K1 = 60 K2 = 3 s = lapply( seq_len(K1), function(i) x^(1+i/30)*cos(x+i) ) series = matrix(nrow=n, ncol=L) for (i in seq_len(n)) series[i,] = s[[I(i,K1)]] + rnorm(L,sd=0.01) getSeries = function(indices) { indices = indices[indices <= n] if (length(indices)>0) series[indices,] else NULL } wf = "haar" getCoefs = function(indices) curvesToCoefs(series[indices,],wf) medoids_K1 = getSeries( clusteringTask(1:n, getCoefs, K1, 75, 4) ) medoids_K2 = computeClusters2(medoids_K1, K2, getSeries, 120) expect_equal(dim(medoids_K1), c(K1,L)) expect_equal(dim(medoids_K2), c(K2,L)) # Not easy to evaluate result: at least we expect it to be better than random selection of # medoids within 1...K1 (among references) distorGood = computeDistortion(series, medoids_K2) for (i in 1:3) expect_lte( distorGood, computeDistortion(series,medoids_K1[sample(1:K1, K2),]) ) })