| 1 | #helper to always have matrices as arg (TODO: put this elsewhere? improve?) |
| 2 | matricize <- function(X) |
| 3 | { |
| 4 | if (!is.matrix(X)) |
| 5 | return (t(as.matrix(X))) |
| 6 | return (X) |
| 7 | } |
| 8 | |
| 9 | require(MASS) |
| 10 | EMGrank = function(Pi, Rho, mini, maxi, X, Y, tau, rank){ |
| 11 | #matrix dimensions |
| 12 | n = dim(X)[1] |
| 13 | p = dim(X)[2] |
| 14 | m = dim(Rho)[2] |
| 15 | k = dim(Rho)[3] |
| 16 | |
| 17 | #init outputs |
| 18 | phi = array(0, dim=c(p,m,k)) |
| 19 | Z = rep(1, n) |
| 20 | # Pi = piInit |
| 21 | LLF = 0 |
| 22 | |
| 23 | #local variables |
| 24 | Phi = array(0, dim=c(p,m,k)) |
| 25 | deltaPhi = c(0) |
| 26 | sumDeltaPhi = 0 |
| 27 | deltaPhiBufferSize = 20 |
| 28 | |
| 29 | #main loop |
| 30 | ite = 1 |
| 31 | while(ite<=mini || (ite<=maxi && sumDeltaPhi>tau)) |
| 32 | { |
| 33 | #M step: Mise à jour de Beta (et donc phi) |
| 34 | for(r in 1:k) |
| 35 | { |
| 36 | Z_indice = seq_len(n)[Z==r] #indices où Z == r |
| 37 | if (length(Z_indice) == 0) |
| 38 | next |
| 39 | #U,S,V = SVD of (t(Xr)Xr)^{-1} * t(Xr) * Yr |
| 40 | s = svd( ginv(crossprod(matricize(X[Z_indice,]))) %*% |
| 41 | crossprod(matricize(X[Z_indice,]),matricize(Y[Z_indice,])) ) |
| 42 | S = s$d |
| 43 | U = s$u |
| 44 | V = s$v |
| 45 | #Set m-rank(r) singular values to zero, and recompose |
| 46 | #best rank(r) approximation of the initial product |
| 47 | if(rank[r] < length(S)) |
| 48 | S[(rank[r]+1):length(S)] = 0 |
| 49 | phi[,,r] = U %*% diag(S) %*% t(V) %*% Rho[,,r] |
| 50 | } |
| 51 | |
| 52 | #Etape E et calcul de LLF |
| 53 | sumLogLLF2 = 0 |
| 54 | for(i in 1:n){ |
| 55 | sumLLF1 = 0 |
| 56 | maxLogGamIR = -Inf |
| 57 | for(r in 1:k){ |
| 58 | dotProduct = tcrossprod(Y[i,]%*%Rho[,,r]-X[i,]%*%phi[,,r]) |
| 59 | logGamIR = log(Pi[r]) + log(det(Rho[,,r])) - 0.5*dotProduct |
| 60 | #Z[i] = index of max (gam[i,]) |
| 61 | if(logGamIR > maxLogGamIR){ |
| 62 | Z[i] = r |
| 63 | maxLogGamIR = logGamIR |
| 64 | } |
| 65 | sumLLF1 = sumLLF1 + exp(logGamIR) / (2*pi)^(m/2) |
| 66 | } |
| 67 | sumLogLLF2 = sumLogLLF2 + log(sumLLF1) |
| 68 | } |
| 69 | |
| 70 | LLF = -1/n * sumLogLLF2 |
| 71 | |
| 72 | #update distance parameter to check algorithm convergence (delta(phi, Phi)) |
| 73 | deltaPhi = c(deltaPhi, max(max(max((abs(phi-Phi))/(1+abs(phi))))) ) |
| 74 | if(length(deltaPhi) > deltaPhiBufferSize){ |
| 75 | l_1 = c(2:length(deltaPhi)) |
| 76 | deltaPhi = deltaPhi[l_1] |
| 77 | } |
| 78 | sumDeltaPhi = sum(abs(deltaPhi)) |
| 79 | |
| 80 | #update other local variables |
| 81 | Phi = phi |
| 82 | ite = ite+1 |
| 83 | } |
| 84 | return(list(phi=phi, LLF=LLF)) |
| 85 | } |