Coupled Matrix and Tensor Factorization

• Coupled matrix and tensor factorization is a framework that jointly decomposes multiway tensors and matrices via shared latent factors, preserving native data structures.
• It enables effective data fusion for applications such as fake news detection, hyperspectral imaging, clinical phenotyping, and recommendation systems.
• Diverse methodologies—including CP, Tucker, PARAFAC2, tensor-ring, and Bayesian approaches—address challenges in scalability, interpretability, and missing data.

Coupled matrix and tensor factorization denotes a family of low-rank data-fusion models in which a tensor and one or more related matrices, or multiple tensors and matrices, are factorized jointly through shared latent factors on common modes. In the simplest setting, a tensor $\mathcal{X}$ and a matrix $Y$ share one index set, so the same factor matrix must simultaneously explain multi-way interactions and side-information; more general formulations allow multiple tensors, arbitrary coupling patterns, shared and private components, linear couplings, and supervised objectives (Choi et al., 2017, Gupta et al., 2018, Khan et al., 2014, Schenker et al., 2022). The central purpose is to combine heterogeneous observations without discarding their native multiway structure, while exploiting common latent structure for prediction, imputation, pattern discovery, recommendation, or representation learning.

1. Definition and data organization

A standard CMTF setting consists of an $N$-way tensor $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$ and one or more matrices $Y^{(r)}$ that share at least one mode with $X$. In one canonical case, a tensor $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$ is coupled with a matrix $\mathbf{Y} \in \mathbb{R}^{I_c \times K}$ through the factor matrix $\mathbf{U}^{(c)}$ on mode $c$, so that the tensor and matrix are explained jointly rather than independently (Choi et al., 2017). A closely related matrix-only problem is coupled matrix factorization, where two matrices $Y$0 and $Y$1 share the row factor $Y$2, with objectives of the form

$Y$3

(Begovic et al., 2024).

The shared mode may correspond to news articles, patients, users, time, firms, spatial locations, or samples, depending on the application. In a fake-news setting, for example, the tensor $Y$4 indexes news, users, and communities, while a content matrix $Y$5 shares the news mode and therefore the same factor matrix $Y$6 (Gupta et al., 2018). In clinical phenotyping, the tensor $Y$7 has modes patients $Y$8 temporal features $Y$9 time, and a matrix $N$0 has modes patients $N$1 static features, again coupled through a patient factor $N$2 (Minoccheri et al., 24 Jun 2025). In hyperspectral super-resolution, the coupling is between a low-resolution hyperspectral image and a high-resolution multispectral image via factors or cores representing the same latent scene (He et al., 2020, Kanatsoulis et al., 2018).

This breadth of data organization is one reason CMTF is best understood as a modeling framework rather than a single decomposition. The literature represented here includes CP/PARAFAC-based models, Tucker-based models, PARAFAC2-based models for ragged tensors, tensor-ring couplings, graph-aware couplings, Bayesian multi-tensor formulations, and deep nonlinear coupled completion models (Choi et al., 2017, Schenker et al., 2022, Khan et al., 2014, Uddin et al., 2021).

2. Canonical factorization models and coupling mechanisms

A widely used formulation couples a CP decomposition of a tensor with a low-rank matrix factorization through a shared factor matrix. In CIMTDetect, the tensor term is

$N$3

the matrix term is

$N$4

and the joint objective is

$N$5

The coupling is implicit: the same $N$6 must explain both the tensor and the matrix, and there are no explicit additional regularization or alignment terms beyond the joint reconstruction loss (Gupta et al., 2018).

Tucker-based CMTF replaces the superdiagonal CP core by a full core tensor. In the sparse Tucker formulation of S3CMTF, the tensor is modeled as $N$7, while the coupled matrix is modeled as $N$8. The paper emphasizes that the Tucker core $N$9 captures inter-relations between factors across modes, in contrast to CP-based methods whose core is hyper-diagonal (Choi et al., 2017). This distinction is methodologically important: CP prioritizes parsimony and identifiability, whereas Tucker increases representational capacity.

PARAFAC2-based CMTF extends the framework to irregular or ragged tensors. A PARAFAC2 slice model takes the form

$X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$0

with slice-specific $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$1 constrained by cross-product invariance, and couples the factor $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$2 to a matrix or another tensor through linear constraints such as

$X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$3

This permits exact sharing, partial sharing, and transformed couplings, while retaining the ability to impose various constraints on all modes (Schenker et al., 2022).

Tensor-ring coupling appears in hyperspectral super-resolution via coupled tensor ring factorization. There, a high-resolution HSI $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$4 is represented by three tensor-ring cores $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$5, and the observed HSI and MSI are obtained by applying spatial or spectral degradation directly to the appropriate cores: $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$6

$X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$7

This structure inherits the simple representation of coupled matrix/CP factorization and flexible low-rank exploration of coupled Tucker factorization (He et al., 2020).

Bayesian multi-tensor factorization generalizes coupled factorization further by modeling arbitrary sets of matrices and tensors, each treated as a view, with factors shared by any subsets of the tensors and factors private to individual tensors (Khan et al., 2014). Joint symmetric tensor factorization for word embeddings offers another special case: the order-2 co-occurrence matrix and order-3 co-occurrence tensor are factorized jointly with the same word factor matrix $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$8, which is explicitly described as related to coupled tensor factorization (Bailey et al., 2017).

Model family	Coupling pattern	Representative sources
CP/PARAFAC + matrix factorization	Shared factor matrix on a common mode	(Gupta et al., 2018)
Tucker + matrix factorization	Shared mode factor with dense core tensor	(Choi et al., 2017, Begovic et al., 2024)
PARAFAC2-based CMTF	Shared or linearly coupled factors for ragged slices	(Schenker et al., 2022)
Tensor-ring coupling	Shared latent scene through degraded cores	(He et al., 2020)
Bayesian multi-tensor factorization	Shared and private factors across arbitrary subsets of views	(Khan et al., 2014)

3. Optimization algorithms and computational regimes

The optimization landscape of CMTF is correspondingly diverse. Tensor-only Tucker models such as CITDetect use HO-SVD / TUCKALS3, a variant of ALS, together with column-wise orthogonality of factor matrices (Gupta et al., 2018). CP-based coupled models such as CIMTDetect use first-order optimization: the tensor gradient is expressed in matricized form with Khatri-Rao products, the matrix gradients are

$X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$9

and the combined gradients are passed to a first-order optimizer; the paper explicitly cites the MATLAB CMTF Toolbox and notes that any first-order optimization algorithm can be employed (Gupta et al., 2018).

Scalability has produced several distinct algorithmic directions. S3CMTF is a sparse Tucker-based method that exploits sparsity, captures inter-relations between factors, and uses lock-free parallel SGD on multi-core shared-memory systems; experimentally it is reported to be $Y^{(r)}$0 times faster and $Y^{(r)}$1 times more accurate than existing methods, with linear scalability on the number of data entries and the number of cores (Choi et al., 2017). Cut-CD, developed for nonnegative CMTF, uses column-wise element selection to prevent frequent gradient updates; the paper states that it is not only more accurate but also more computationally efficient than existing algorithms in approximating the tensor as well as in identifying the underlying nature of factors (Balasubramaniam et al., 2020).

A different line of work replaces iterative coupling solvers by direct low-rank reduction. For coupled matrix factorization,

$Y^{(r)}$2

is equivalent to

$Y^{(r)}$3

so the solution can be obtained from the best rank-$Y^{(r)}$4 approximation of the concatenated matrix. The same idea extends to Tucker- and CP-based CMTF via mode-1 matricization, after which randomized RSVD-, RSI-, and RBKI-based algorithms reduce computational cost while preserving accuracy (Begovic et al., 2024). This directly challenges the assumption that coupled decompositions must be solved iteratively.

In constrained and structured settings, ADMM-based schemes are prominent. CGTF employs an ADMM algorithm with closed-form updates for graph-coupled tensor factorization (Ioannidis et al., 2018), and PARAFAC2-based CMTF uses an AO-ADMM framework to impose various constraints on all modes and linear couplings (Schenker et al., 2022). Supervised CMTF in ulcerative colitis adopts all-at-once optimization in a deep learning framework, with Adam for decomposition parameters, SGD for classifier parameters, projected gradient descent for nonnegativity, and proximal gradients for $Y^{(r)}$5-based sparsity (Minoccheri et al., 24 Jun 2025). Deep nonlinear coupled completion in MLCTR likewise uses SGD, but moves non-linearity into the factor matrices through multi-layer low-rank matrix-factorization blocks with interleaved transfer functions and bypass connections (Uddin et al., 2021).

Taken together, these results indicate that computational strategy is not secondary in CMTF: it often determines whether coupling remains feasible at the sparsity, dimensionality, and heterogeneity scales for which the framework is intended.

4. Constraints, identifiability, and structural variants

Constraint design is central to coupled factorization because the coupling alone does not determine interpretability, uniqueness, or robustness. The literature here includes nonnegativity, orthogonality, sparsity, masking for missing data, graph structure, bias terms, and supervision. In N-CMTF, all factors are constrained to be elementwise nonnegative, which the Cut-CD work explicitly motivates for identifying latent patterns, prediction, and recommendation (Balasubramaniam et al., 2020). In CITDetect, the factor matrices are column-wise orthogonal as in standard HOSVD (Gupta et al., 2018). In SCMTF, all factors and weights are constrained to be nonnegative, the tensor loss is masked by $Y^{(r)}$6 to ignore missing entries, and an $Y^{(r)}$7 penalty

$Y^{(r)}$8

encourages sparse phenotypes (Minoccheri et al., 24 Jun 2025).

Graph-aware coupling changes the matrix model itself. In CGTF, each graph $Y^{(r)}$9 is not modeled as a generic low-rank matrix but via diagonally scaled symmetric nonnegative matrix factorization,

$X$0

with $X$1 shared with the CP tensor model. This explicitly embeds community structure into the coupled factorization and supports both recommendation and community detection when some links in the graphs are missing (Ioannidis et al., 2018).

Bias terms can be structurally necessary rather than cosmetic. In SCMTF, patient and feature bias tensors are added to the CP reconstruction to account for subjectivity in patient-reported outcomes and the fact that lab values are continuous and have strictly positive baselines whereas PROs are discrete and significantly more sparse (Minoccheri et al., 24 Jun 2025). This paper also reports that all-at-once optimization with bias outperforms ALS-style baselines, and that the best model predicts changes in medication 8 and 20 months in the future with AUCs of $X$2 and $X$3 on the test set (Minoccheri et al., 24 Jun 2025).

Identifiability is model-dependent. In hyperspectral super-resolution, the coupled tensor factorization approach is stated to guarantee the identifiability of the SRI under mild and realistic conditions, and to work with little knowledge of the degradation operators (Kanatsoulis et al., 2018). In coupled tensor-ring factorization, Theorem 1 establishes that

$X$4

which motivates a spectral nuclear norm regularization on the spectral core in NCTRF (He et al., 2020). Bayesian MTF adds a different kind of structure: factors can be shared by any subsets of the tensors or private to individual tensors, and rMTF relaxes the strict CP/trilinear assumption into a continuum between CP and Tucker-1-like models (Khan et al., 2014).

A persistent misconception is that coupling merely means setting two factor matrices equal. The literature here is broader. Coupling may be implicit through a shared factor in a joint reconstruction loss (Gupta et al., 2018), explicit through linear constraints involving a latent variable $X$5 (Schenker et al., 2022), structural through graph factorizations (Ioannidis et al., 2018), or probabilistic through priors that decide which factors are shared or private (Khan et al., 2014). This suggests that “coupling” is best understood as a design principle for latent-variable sharing rather than a single algebraic constraint.

5. Representative applications

Fake-news detection provides a compact example of how CMTF merges relational and semantic evidence. CIMTDetect constructs a news–user–community tensor from social sharing and an article-content matrix from bigram counts, couples them through the news factor $X$6, and reports that on both BuzzFeed and PolitiFact datasets CIMTDetect achieves the highest F1-score among all methods, while the learned embeddings also improve News Cohort Analysis and Collaborative News Recommendation (Gupta et al., 2018).

Computational phenotyping in ulcerative colitis shows the same principle in a clinical setting. SCMTF jointly factorizes a temporal tensor of labs and patient-reported outcomes with a static matrix of demographics and medication history, shares a patient factor across both, adds a supervised classifier on the patient memberships, and derives interpretable phenotypes containing static features and temporal features, including their temporal patterns (Minoccheri et al., 24 Jun 2025). The reported best test AUCs are $X$7 for year-2 medication persistence and $X$8 for year-3 persistence (Minoccheri et al., 24 Jun 2025).

Hyperspectral super-resolution is a major imaging application. Coupled matrix, CP, Tucker, and tensor-ring formulations all appear in the provided literature. The CTRF model simultaneously learns a high spectral resolution core tensor from the HSI and high spatial resolution core tensors from the MSI, reconstructing the HR-HSI via tensor ring representation; the paper reports that NCTRF compares favorably with previous matrix/tensor and deep learning methods (He et al., 2020). The coupled tensor factorization approach of STEREO addresses the same fusion problem from a tensor-identifiability perspective, emphasizing that matrix-based fusion may lose structural information and that the tensor model guarantees identifiability of the SRI under mild and realistic conditions (Kanatsoulis et al., 2018).

Recommendation and community analysis are longstanding CMTF domains. CGTF treats side-information matrices explicitly as graphs, uses shared nonnegative factors for both tensor and graph terms, and is reported to be successful both for recommendations and for community detection even when some links in the graphs are missing (Ioannidis et al., 2018). S3CMTF was applied to Yelp recommendation tensor data coupled with three additional matrices to discover interesting properties (Choi et al., 2017). In financial completion, MLCTR couples an EPS tensor with a fundamentals tensor through shared time and firm factors, and its abstract states that the model is highly efficient for imputing missing values in the EPS data while outperforming approaches with non-linearity in the phase of reconstructing tensors from factor matrices (Uddin et al., 2021).

Multimodal neuroimaging supplies a supervised use-case. C-STM builds a maximal-margin classifier on latent factors jointly estimated from ACMTF, combining individual and shared latent factors with multiple kernels; in simultaneous EEG-fMRI analysis it reports accuracies of $X$9 and $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$0 and corresponding AUCs of $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$1 and $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$2 for auditory and visual tasks, respectively (Peide et al., 2022).

Language and topic models illustrate that coupled factorization is not limited to numeric sensing data. Joint symmetric factorization of a word co-occurrence matrix and a third-order co-occurrence tensor shares a single embedding matrix $X \in \mathbb{R}^{I_1 \times \cdots \times I_N}$3, and the authors state that embeddings based on tensor factorization outperform existing matrix-based methods on an evaluation of higher-order relations (Bailey et al., 2017). A constrained coupled matrix-tensor factorization has also been proposed for learning time-evolving and emerging topics, where topics are characterized not only by temporal evolution but also by level of difficulty inferred from contributor expertise; the abstract reports implications for automatic curriculum design (Bahargam et al., 2018).

6. Methodological tensions and current directions

Several methodological tensions recur across this literature. One is the trade-off between expressive flexibility and structural discipline. CP-based couplings are simple and often more identifiable; Tucker-based models can capture inter-relations between factors; PARAFAC2 accommodates irregular tensors; tensor-ring coupling separately exploits classwise low-rank structure in HSR; and Bayesian rMTF deliberately relaxes strict trilinearity toward Tucker-1-like behavior (Choi et al., 2017, Schenker et al., 2022, He et al., 2020, Khan et al., 2014). This suggests that model choice should follow the geometry of the data—shared modes, raggedness, degradation operators, or need for private factors—rather than the existence of a generic “best” decomposition.

A second tension concerns iterative versus direct solvers. Classical CMTF work is dominated by ALS, first-order optimization, ADMM, and SGD, yet randomized coupled decompositions show that both CMF and certain Tucker- and CP-based CMTF problems can be reduced to low-rank approximation of concatenated matricizations and solved by direct or randomized SVD (Begovic et al., 2024). That result does not eliminate iterative methods; rather, it narrows the class of situations in which iteration is algorithmically necessary.

A third tension is between unsupervised reconstruction and task-driven latent structure. Most classical CMTF formulations minimize reconstruction error, whereas SCMTF and C-STM place prediction directly into or immediately above the factorization, and MLCTR uses deep nonlinear embedding learning for coupled completion (Minoccheri et al., 24 Jun 2025, Peide et al., 2022, Uddin et al., 2021). A plausible implication is that coupling is increasingly being used as a representation-learning substrate, not merely as a decomposition for exploratory analysis.

Finally, missingness, heterogeneity, and structure-specific priors remain active frontiers. The provided papers include masking for missing tensor entries (Minoccheri et al., 24 Jun 2025), graph-specific side models (Ioannidis et al., 2018), randomization for very large matrices and tensors (Begovic et al., 2024), deep nonlinear factor learning (Uddin et al., 2021), and generalized Bayesian multi-tensor formulations that allow factors shared by any subsets of datasets (Khan et al., 2014). The combined picture is of a field that has moved well beyond the original “one tensor plus one matrix with a shared mode” template while retaining that template as its algebraic core.