Multilinear Discriminant Analysis

• Multilinear Discriminant Analysis (MDA) is a tensor-based supervised dimension reduction technique that preserves the inherent structure of high-order data to enhance class separability.
• It employs mode-specific projections and generalized eigenproblems, along with transform-domain optimizations, to effectively mitigate the curse of dimensionality.
• Advanced variants like HOMLDA and BTTDA offer improved computational efficiency and interpretability, demonstrating strong performance in applications such as computer vision and neuroimaging.

Multilinear Discriminant Analysis (MDA) is a class of methods for supervised dimension reduction where the input data are higher-order arrays (tensors) and the objective is to project samples to a low-dimensional multilinear subspace that maximally separates classes. Unlike classical vector-based Linear Discriminant Analysis (LDA), which requires vectorization and often destroys the inherent mode-wise structure of multidimensional data (e.g., images, videos, neuroimaging signals), MDA preserves the tensorial relationships and greatly mitigates the curse of dimensionality. Central to MDA is the formulation of discriminative projections—either as sets of mode-specific matrices or stacked tensor operators—that optimize a criterion of inter-class separability to intra-class compactness over tensor representations.

1. Mathematical Foundations

Let $$\mathcal{X}_m\in\mathbb{R}^{I_1\times I_2\times\cdots\times I_N}$$, $m=1,\dots,M$ denote $M$ samples as $N$-th order tensors across $C$ classes. The canonical MDA objective extends the Fisher criterion to tensors by constructing mode-specific projections $\{u^{(n)}\}_{n=1}^N$, each $u^{(n)}\in\mathbb{R}^{I_n}$, to define an elementary multilinear projection (EMP) to a scalar:

$$y_m = \mathcal{X}_m \times_1 (u^{(1)})^T \times_2 (u^{(2)})^T \cdots \times_N (u^{(N)})^T \in \mathbb{R}.$$

Inter-class and intra-class scatter for the scalar projections are

$$S_B = \sum_{c=1}^C N_c (\bar{y}_c - \bar{y})^2, \qquad S_W = \sum_{c=1}^C \sum_{m\in\mathcal{C}_c} (y_m - \bar{y}_c)^2,$$

where $\bar{y}_c$ is the class-$c$ mean and $\bar{y}$ the global mean. The Fisher criterion is then

$$J(u^{(1)},\dots,u^{(N)}) = \frac{S_B}{S_W} \quad \text{subject to}\;\|u^{(n)}\|=1\;\forall n.$$

A full MLDA extracts $P$ such EMPs and maximizes $\sum_{p=1}^P S_B^{(p)} / S_W^{(p)}$ (Zeng et al., 2014). For order-$N$ tensors, classical approaches use alternating optimization, mode-wise eigenproblems, and tensor-matrix products (Dufrenois et al., 2022, Tran et al., 2017).

2. Algorithmic Strategies and Tensor-Specific Optimizations

The non-convexity of the joint objective in MLDA leads to alternating, mode-by-mode updates: each mode-$n$ vector or matrix is optimized via a rank-1 Fisher-LDA (or generalized eigenproblem) on unfolded within- and between-class scatter matrices while holding other mode projections fixed. Formally, for each mode: $$S_B^{(n)} u^{(n)} = \lambda S_W^{(n)} u^{(n)}, \quad \|u^{(n)}\|=1,$$ where mode-$n$ scatter matrices are constructed via mode-$n$ unfolding and multi-mode contraction with fixed factors (Zeng et al., 2014, Tran et al., 2017).

Transform-domain generalizations, exploiting invertible transforms $L$ (e.g., FFT, DCT), further decouple the optimization: MDA is reframed such that per-frontal slice generalized eigenproblems can be solved independently and efficiently in the transform domain, and combined via inverse transformations to yield the final projector (Dufrenois et al., 2022, Ozdemir et al., 2022). For an arbitrary transform $L$, let $L(\mathcal{X}) = \widetilde{\mathcal{X}}$, with the $L$-product defined by per-slice multiplication. This fully tensorial framework allows for robust, block-diagonal decompositions and avoids repeated tensor-matrizations and mode-wise back-and-forth.

Block-term decompositions, such as in Block-Term Tensor Discriminant Analysis (BTTDA), extend the parameterization via sums of multiple supervised Tucker blocks, each optimizing a Fisher ratio and extracted in a deflationary fashion (Kerchove et al., 6 Nov 2025). BTTDA thus yields both higher expressivity and parameter efficiency.

3. Generalizations and Recent Frameworks

Several methodological innovations generalize the core MDA framework:

• Transform-Domain MDA: Reformulates MDA using the $L$-product, enabling parallelization over the third (or higher) tensor mode by solving smaller, independent generalized eigenproblems, e.g., c-TDA (DCT-based) and t-TDA (FFT-based) (Dufrenois et al., 2022). The transform allows for data-dependent adaptation and increased numerical stability.
• High-Order MDA (HOMLDA): Lifts generalized eigenanalysis directly to order-$n$ tensors via the $*_L$ tensor product and tensor-eigen decomposition ($t\texttt{-eig}_L$). HOMLDA provides a non-iterative approach, mimicking LDA’s operator structure at the tensor level (Ozdemir et al., 2022).
• Robust HOMLDA (RHOMLDA): When within-class scatter tensors are ill-conditioned, RHOMLDA applies slice-by-slice spectral regularization to stabilize the inversion and preserve discriminant structure, leading to substantial empirical gains (Ozdemir et al., 2022).
• Multilinear Class-Specific Discriminant Analysis (MCSDA): Forms class-specific discriminant subspaces by maximizing the out-of-class to in-class scatter ratio for each target class (one-vs-rest), leveraging the tensor structure for scalability and spatial/temporal coherence (Tran et al., 2017).
• Block-Term Tensor Discriminant Analysis (BTTDA): Implements dimensionality reduction via additive block-supervised Tucker decompositions, accommodating greater flexibility and interpretability of discriminant directions (Kerchove et al., 6 Nov 2025).

4. Integration with Deep Architectures: MLDANet

Multilinear Discriminant Analysis can be embedded within convolution-style networks for tensor object classification, as exemplified by MLDANet (Zeng et al., 2014). MLDANet operates as follows:

• Stage-1 Encoder: MLDA is run on local tensor patches to learn $L_1$ elementary multilinear projections (EMPs), acting as discriminant convolutional filters.
• Stage-2 Decoder: A standard LDA is applied on the outputs of the first stage, deriving $L_2$ linear filters for further projection of feature maps.
• Pooling: Binary hashing and block-wise histograms are used to aggregate the filtered features, forming a final, compact feature vector for classification.

Empirical results on the UCF11 action recognition database demonstrate that two-stage MLDANet achieves superior accuracy (78.93%) compared to single-stage variants and other tensor LDA baselines (e.g., PCANet, LDANet, MPCA+LDA, and standalone MLDA), confirming the necessity of preserving local multilinear structure before projection to vector spaces (Zeng et al., 2014).

5. Empirical Performance and Application Domains

MDA and its generalizations have been applied extensively in computer vision (face recognition, gait analysis, action recognition), finance (order book-based stock prediction), and brain-computer interfacing (EEG/ERP/MI decoding). Results include:

• Transform-domain c-TDA/t-TDA achieve accuracy improvements over vector LDA/Fisherfaces (e.g., DIV: 90.6% vs 83.1%; AR: 95.6% vs 92.3%) and markedly faster training via parallel slicewise decompositions (Dufrenois et al., 2022).
• HOMLDA/RHOMLDA outperform Tucker-based CMDA and DGTDA, particularly when within-class covariance is ill-conditioned (e.g., UTD-MHAD: RHOMLDA-DCT 97.45% vs DGTDA 94.91%) (Ozdemir et al., 2022).
• BTTDA yields highest ROC-AUC in ERP decoding (91.25% vs HODA 88.89% and PARAFACDA 90.94%) and is statistically superior by Wilcoxon tests (Kerchove et al., 6 Nov 2025).
• In class-specific tasks, MCSDA delivers mAP and F1-scores superior to both global MDA and Bag-of-Words/Neural BoW, while retaining substantial computational efficiency (Tran et al., 2017).

The following table summarizes representative empirical results across core MDA variants:

Method	Benchmark Task	Reported Accuracy / AUC
MLDANet-2	UCF11 video action	78.93%
c-TDA	DIV image	90.6%
RHOMLDA-DCT	UTD-MHAD action	97.45%
BTTDA	ERP EEG ROC-AUC	91.25%
MCSDA	Order book F1	46.7%

6. Theoretical and Practical Considerations

MDA’s principal advantage lies in exploiting the tensor structure to reduce the number of free parameters, mitigate overfitting, and avoid the small-sample-size problem inherent in high-dimensional vectorizations. Alternating mode-wise optimization, however, is not guaranteed to reach the global optimum; non-iterative methods via tensor eigendecomposition or transform-domain decoupling address this limitation in specific settings (Dufrenois et al., 2022, Ozdemir et al., 2022).

Choice of tensor decomposition and product (Tucker, PARAFAC, $L$-product, block-term) impacts both expressivity and interpretability. Block-term models such as BTTDA facilitate physiologically interpretable subspaces, particularly important in neuroimaging, by isolating distinct sources of discriminant variance (Kerchove et al., 6 Nov 2025).

Regularization is crucial: for ill-conditioned within-class scatter tensors, direct inversion deteriorates results. Slicewise thresholding and eigenvalue correction (as in RHOMLDA) provide robust alternatives, notably beneficial in dynamic gesture and action datasets with high intra-class variability (Ozdemir et al., 2022).

7. Extensions and Ongoing Directions

Research continues toward generalized tensor–tensor algebraic frameworks, scalable optimization (transform domain, parallelism), class-specific and one-vs-rest discriminants, and end-to-end deep learning integrations. Notable trajectories include kernelized/nonlinear MDA variants, online/incremental tensor discriminant updates, and domain-adaptive transformation selection (e.g., learning optimal $L$ for data).

BTTDA’s iterative, deflationary supervised block extraction exemplifies recent progress toward interpretable, high-capacity multilinear models, with applications broadening beyond EEG/BCI to multi-modal sensing, spatiotemporal analytics, and structured output prediction.

---

References:

• "Tensor object classification via multilinear discriminant analysis network" (Zeng et al., 2014)
• "Multilinear Discriminant Analysis using a new family of tensor-tensor products" (Dufrenois et al., 2022)
• "High-Order Multilinear Discriminant Analysis via Order-$n$ Tensor Eigendecomposition" (Ozdemir et al., 2022)
• "Multilinear Class-Specific Discriminant Analysis" (Tran et al., 2017)
• "BTTDA: Block-Term Tensor Discriminant Analysis for Brain-Computer Interfacing" (Kerchove et al., 6 Nov 2025)