Skip to content

Compare methods

Compare source explanations, strengths, limitations and references for up to four methods.

8 results
Criterion Matrix Completion MICE Non-negative Matrix Factorization
Overview
  • Matrix Completion is a technique for recovering a low-rank matrix from a small, possibly random subset of its entries. Introduced by Emmanuel Candès and Benjamin Recht in 2009, it reformulates the problem as nuclear norm minimization — a convex surrogate for rank minimization — and provides theoretical guarantees that exact recovery is achievable when entries are observed uniformly at random and the matrix satisfies an incoherence condition.
  • Multivariate Imputation by Chained Equations (MICE) is an iterative procedure for handling missing data in multivariate datasets. Introduced by Stef van Buuren and Karin Groothuis-Oudshoorn through the R package mice (2011), the algorithm fills each missing variable using a separate regression model conditioned on all other variables, cycling through variables repeatedly until the imputed values converge. The result is m completed datasets that are analysed separately and combined using Rubin's rules.
  • Non-negative Matrix Factorization (NMF) is a family of algorithms, introduced by Lee and Seung in their landmark 1999 Nature paper, that decomposes a non-negative data matrix V into the product of two lower-rank non-negative matrices W (basis components) and H (encoding coefficients). Unlike PCA or SVD, the non-negativity constraint forces the algorithm to learn strictly additive, parts-based representations, making the factors directly interpretable as building blocks of the original data.
Use and interpretation
  • Use matrix completion when data can be organized as a matrix with systematically missing entries and there is reason to believe the true signal has low rank — for example, in recommender systems, genomics, social network analysis, and image recovery. Key assumptions are that entries are missing at random and the matrix satisfies an incoherence condition. It is not appropriate for matrices with structured or non-random missingness, very high-rank structure, or when the low-rank assumption is unjustified. Alternatives include MICE imputation for general missing data and NMF when non-negativity is required.
  • MICE is appropriate when data are missing at random (MAR) and missingness occurs across multiple variables simultaneously. It suits continuous, binary, ordinal, and count outcomes through its flexible per-variable model specification. It is less suitable for data missing not at random (MNAR) without auxiliary modelling, for very high-dimensional datasets where the imputation model becomes unwieldy, or when missingness rates exceed 50% in key variables. Single imputation or complete-case analysis may suffice under MCAR with low missingness rates.
  • NMF is appropriate when the data are non-negative by nature — pixel intensities, word counts, spectral amplitudes, gene expression levels — and when interpretability of the latent components matters. It is well-suited to topic modelling of text corpora, hyperspectral image unmixing, audio source separation, and bioinformatics applications such as identifying gene expression signatures. A key assumption is that the observed data are plausibly generated as non-negative combinations of underlying non-negative patterns; if the data contain negative values, NMF is not applicable without preprocessing. The number of components k must be specified in advance, so model selection (e.g., via reconstruction error, stability across runs, or domain knowledge) is part of the workflow.
Output
  • Completed low-rank matrix
Strengths
  • Provably exact recovery under mild conditions with a polynomial-time convex program
  • Handles very high rates of missingness when the matrix is truly low-rank
  • No prior knowledge of rank is required; the nuclear norm minimization implicitly finds it
  • Strong theoretical guarantees rooted in random matrix theory and convex analysis
  • Handles mixed variable types (continuous, binary, ordinal, count) via tailored per-column models
  • Properly propagates uncertainty through multiple imputed datasets and Rubin's pooling rules
  • Flexible predictor specification allows domain knowledge and auxiliary variables to improve imputation quality
  • Well-validated R implementation (mice package) with extensive diagnostics and active maintenance
  • Non-negativity produces parts-based, additive representations that are directly interpretable without sign conventions.
  • Multiplicative update rules are simple to implement, preserve non-negativity automatically, and have guaranteed monotone convergence.
  • Competitive with or superior to PCA and VQ for learning parts of objects such as facial features and semantic topics in text.
  • Well-established theoretical and algorithmic foundations, with decades of extensions covering sparsity, robustness, and probabilistic formulations.
  • Applicable to a wide range of domains wherever data are inherently non-negative: images, text, audio, spectral data, and genomics.
Limitations
  • Computational cost of semidefinite programming scales poorly with matrix dimensions
  • Incoherence assumption may fail in real datasets with structured missingness or outliers
  • Performance degrades when the true rank is not much smaller than the matrix dimensions
  • Nuclear norm minimization can be sensitive to noise; robust variants add additional complexity
  • Assumes data are missing at random (MAR); violations require sensitivity analyses or pattern-mixture extensions
  • Computational cost grows with the number of variables, iterations, and imputations (m)
  • Convergence of the Gibbs-like sampler is not mathematically guaranteed for incompatible conditional models
  • Model specification (predictor selection, imputation method per variable) requires careful analyst judgment
  • Non-convex optimisation means that results depend on initialisation and convergence to a global minimum is not guaranteed.
  • The number of components k must be chosen prior to fitting; there is no automatic selection and model comparison can be computationally expensive.
  • Not applicable to data containing negative values without domain-specific preprocessing that may distort the analysis.
  • Convergence of multiplicative update rules can be slow for large matrices; projected gradient and alternating least squares variants are often faster in practice.
  • Uniqueness of the factorisation is generally not guaranteed unless additional constraints such as sparsity or orthogonality are imposed.
Common pitfalls
  • Assuming uniform random missingness when real observations are user-selected or biased (e.g., only highly rated items are rated), which violates the theoretical sampling model
  • Neglecting to check the incoherence condition, leading to poor recovery even with sufficient observations
  • Using nuclear norm minimization without scaling or regularization in the noisy setting, where the exact-recovery formulation is inappropriate
  • Conflating matrix completion with matrix factorization heuristics: alternating least squares may converge to local minima and lacks the recovery guarantees of nuclear norm minimization
  • Using too few imputed datasets (m < 5) underestimates standard errors and inflates false-positive rates
  • Including the outcome variable in the imputation model is essential; omitting it biases regression coefficients toward null
  • Ignoring interaction terms or non-linear relationships in the imputation model can distort downstream analyses
  • Failing to check convergence via trace plots may lead to pooling results from non-converged chains
  • Fitting NMF only once with a single random initialisation and reporting the result without checking sensitivity to starting conditions.
  • Treating k as a free parameter to be tuned purely by reconstruction error, without considering stability of the components across runs.
  • Applying NMF to data that are not non-negative, or normalising data in a way that introduces negative values before factorisation.
  • Ignoring scale ambiguity: W and H can be rescaled inversely without changing WH, so comparing component magnitudes across runs requires normalisation.
  • Reporting NMF components as interpretable without verifying that the reconstruction error is acceptably low.
Introduced by
  • Emmanuel Candès & Benjamin Recht
  • Stef van Buuren & Karin Groothuis-Oudshoorn
  • Lee, D. D. & Seung, H. S.
Year
  • 2009
  • 2011
  • 1999