SVD Imputation

Страница в процессе перевода.

Often matrices and n-dimensional arrays with missing values can be imputed via a low rank approximation. Impute.jl provides one such method using a single value decomposition. The general idea is to:

Fill the missing values with some rough approximates (e.g., mean, median, rand)
Reconstruct this "completed" matrix with a low rank SVD approximation (i.e., k largest singular values)
Replace our initial estimates with the reconstructed values
Repeat steps 1-3 until convergence (update difference is below a tolerance)

To demonstrate how this is useful lets load a reduced MNIST dataset. We’ll want both the completed dataset and another dataset with 35% of the values set to -1.0 (indicating missingness).

TODO: Update example with more a realistic dataset like some microarray data

using Distances, Impute, Plots, Statistics, DataFrames, CSV
mnist = Impute.dataset("test/matrix/mnist");
completed, incomplete = mnist[0.0], mnist[0.25];

Alright, before we get started lets have a look at what our incomplete data looks like:

heatmap(incomplete; color=:greys);

Okay, so as we’d expect there’s a reasonable bit of structure we can exploit. So how does the svd method compare against other common, yet simpler, methods?

data = Impute.declaremissings(incomplete; values=-1.0)

# NOTE: SVD performance is almost identical regardless of the `init` setting.
imputors = [
    "0.5" => Impute.Replace(; values=0.5),
    "median" => Impute.Substitute(),
    "svd" => Impute.SVD(; tol=1e-2),
]

results = map(last.(imputors)) do imp
    r = Impute.impute(data, imp; dims=:)
    return nrmsd(completed, r)
end

bar(first.(imputors), results);