EMPCA notes"

In nipals: Principal Components Analysis using NIPALS or Weighted EMPCA, with Gram-Schmidt Orthogonalization

Compare nipals/empca with no missing Compare nipals/empca with missing, equal weight Compare nipals/empca with missing, unequal weight

Complete data

# Python - Coeff (scores)
[[-2.809  0.097  0.244  0.050]
 [-1.834  0.286  0.010 -0.135]
 [-0.809  0.963 -0.341  0.078]
 [-0.155 -1.129  0.548  0.026]
 [0.707  -0.723 -0.736 -0.024]
 [1.830  -0.290 -0.157  0.030]
 [3.070   0.796  0.431 -0.026]]

# m1e <- empca(x=B1, w=B1wt, ncomp=4)
# Un-sweep the eigenvalues to compare to python results
# R round( sweep( m1e$scores, 2, m1e$eig, "*"), 3)
      PC1    PC2    PC3    PC4
G1 -2.809  0.097 -0.244  0.050
G2 -1.834  0.286 -0.010 -0.135
G3 -0.809  0.963  0.341  0.078
G4 -0.155 -1.129 -0.548  0.026
G5  0.707 -0.723  0.736 -0.024
G6  1.830 -0.290  0.157  0.030
G7  3.070  0.796 -0.431 -0.026

# Matlab - P (scores)
  0.5590   0.0517   0.2210   0.2910
  0.3650   0.1520   0.0095  -0.7840
  0.1610   0.5120  -0.3080   0.4530
  0.0309  -0.6010   0.4950   0.1510
 -0.1410  -0.3850  -0.6640  -0.1380
 -0.3650  -0.1540  -0.1420   0.1760
-0.6110   0.4230   0.3890  -0.1490

# R round(m1e$scores, 3)
      PC1    PC2    PC3    PC4
G1 -0.559 -0.052  0.221 -0.291
G2 -0.365 -0.152  0.009  0.784
G3 -0.161 -0.512 -0.308 -0.453
G4 -0.031  0.601  0.495 -0.151
G5  0.141  0.385 -0.664  0.138
G6  0.365  0.154 -0.142 -0.176
G7  0.611 -0.423  0.389  0.149

Missing data

# Python with initial Identity matrix

[[2.791 0.125 0.325 -0.035]
 [1.528 -0.989 -0.211 0.172]
 [0.990 -0.651 -0.117 -0.186]
 [0.159 1.463 0.530 0.020]
 [-0.628 0.862 -0.730 -0.032]
 [-1.738 0.406 -0.139 -0.071]
 [-2.917 -0.712 0.520 -0.047]]

Eigvec (loadings)
[[-0.309 -0.839 -0.298 0.300]
 [-0.502 0.014 0.154 -0.615]
 [-0.470 -0.086 0.766 0.219]
 [-0.441 0.521 -0.236 0.615]
 [-0.487 0.128 -0.496 -0.326]]


# R 

R> m2e <- empca(x=B2, w=B2wt, ncomp=4, seed=NULL)
# # Un-sweep the eigenvalues to compare to python results
R> round( sweep( m2e$scores, 2, m2e$eig, "*"), 3)
      PC1    PC2    PC3    PC4
G1 -2.791  0.216 -0.356  0.066
G2 -1.528 -0.942  0.187 -0.150
G3 -0.990 -0.620  0.101  0.207
G4 -0.159  1.472 -0.522 -0.032
G5  0.628  0.844  0.744  0.021
G6  1.738  0.351  0.161  0.050
G7  2.917 -0.808 -0.493  0.019
R> round( m2e$loadings, 3)
     PC1    PC2    PC3    PC4
E1 0.309 -0.839  0.298 -0.300
E2 0.502  0.014 -0.154  0.615
E3 0.470 -0.086 -0.766 -0.219
E4 0.441  0.521  0.236 -0.615
E5 0.487  0.128  0.496  0.326

Python

Python code by @bailey2012principal, retrieved 1 Mar 2019 from https://github.com/sbailey/empca .

The Python code is difficult to read in places for a person [like me] not well-versed with Python. Three examples:

1. It is not clear what values k takes in for k in range(self.nvec).
2. Gram-Schmidt orthogonalization is accomplished with a pair of nested for loops instead of a function.
3. The Model class structure makes it a bit tricky to figure out what objects have actually been modified inside a function.

The Python code iterates these two EM steps:

1. Calculate the coefficient matrix C.
2. Calculate ALL ncomp principal components P simultaneously (iterate each to convergence). Orthogonalize P.

For a complete-data problem, python and R give similar results. Note the Coeff matrix in python does NOT have eigenvalues swept out of the columns.

For the missing-data problem, the python results are somewhat different from R.

Matlab

Matlab code by Vicente Parot, retrieved 1 Mar 2019 from https://www.mathworks.com/matlabcentral/fileexchange/45353-empca.

The Matlab code feels similar to R.

The Matlab code calculates principal components sequentially, one at time. For each principal component, the algorithm iterates between these two steps:

1. Calculate C[,h]
2. Calculate P[,h]

While this is a type of EM algorithm, it is not the algorithm described by Bailey (2012) and is considered further.

nipals documentation built on Sept. 16, 2021, 1:07 a.m.