Partial Modal Tensor Regression: A New Robust Modeling Approach for High-Dimensional Tensor Data

Tensor data are increasingly common in modern scientific research, ranging from neuroimaging to recommender systems, where multi-dimensional arrays carry rich structural information. Characterizing the relationship between tensor predictors and responses in high-dimensional statistics while preserving the intrinsic structure of tensors has long been a challenge. Most existing tensor regression methods focus on conditional means or conditional quantiles, but their performance degrades when the response distribution is asymmetric, heavy-tailed, or contaminated by outliers. A recent paper proposes Partial Modal Tensor Regression (PMTR), offering a robust and interpretable alternative to this problem.

Why Do We Need Modal Regression?

In regression analysis, the conditional mean is sensitive to outliers, and conditional quantiles, while robust, only characterize specific positions of the distribution. The conditional mode instead focuses on the most probable value in the response distribution—the peak of the probability density function. When the distribution is skewed or heavy-tailed, the mode often better represents the "typical" response than the mean and is insensitive to outliers. Introducing the mode into tensor regression can achieve robustness while preserving tensor structure.

Conditional mode refers to the peak point of the conditional probability density function of the response variable given the predictors. Unlike the conditional mean (the center of gravity of the distribution) and conditional quantiles (specific cut points of cumulative probability), the mode characterizes the "most probable value," i.e., the location of highest density. For unimodal distributions, the mode is insensitive to skewness and heavy tails because the peak location is determined primarily by local density, and outliers do not pull the mode the way they pull the mean. In multimodal distributions, one must distinguish between the global mode and local modes, where the global mode corresponds to the peak with the highest density. Estimation in modal regression is typically achieved via kernel density estimation or empirical likelihood methods, and its statistical properties are more complex than those of mean regression, with convergence rates depending on the curvature of the density near the peak. It is precisely this locality that makes the mode a natural target for robust regression.

The Core Idea of PMTR

The PMTR method proceeds in two steps. First, it constructs a partial modal tensor covariance, using information from the conditional mode to guide tensor dimension reduction. The key here is that the construction of the covariance matrix incorporates modal information, thereby selecting the tensor subspace most relevant to the response mode. Subsequently, a modal tensor regression model is built in the reduced tensor space.

This strategy avoids the curse of dimensionality caused by direct vectorization. Vectorizing tensors destroys the multi-dimensional structure of the data and leads to an explosion in the number of parameters, whereas PMTR preserves the intrinsic structural information of tensor predictors through dimension reduction while significantly reducing the dimensionality burden.

Tensor dimension reduction is the key step that distinguishes PMTR from vectorization-based methods. If a $p_1 \times p_2 \times \cdots \times p_K$ tensor is directly flattened into a vector, its dimension is $\prod_k p_k$, and the number of parameters grows exponentially with the number of modes—this is the "curse of dimensionality" in tensor regression. The core assumption of tensor dimension reduction is that the tensor predictor has a low-rank or low-dimensional subspace structure, i.e., only a few directions along each mode are relevant to the response. Common dimension reduction techniques include Tucker decomposition (independent reduction per mode) and CP decomposition (shared rank across modes). PMTR uses the partial modal tensor covariance to select the mode directions most relevant to the response mode, which is equivalent to supervised feature extraction while preserving the multi-dimensional tensor structure, thereby reducing the effective number of parameters from exponential order to the order of the sum of the dimensions across modes.

Theoretical Properties and Dimension Selection

The paper establishes consistency and asymptotic properties of the estimator under appropriate regularity conditions, providing theoretical guarantees for the method. In addition, the authors adopt a data-driven dimension selection procedure to determine the effective dimension along each tensor mode. This means that in practice no manual tuning is needed—the algorithm can adaptively select the reduced dimensions, enhancing the method's practicality.

Data-driven dimension selection in tensor regression typically employs BIC-type criteria, cross-validation, or eigenvalue ratio methods. Since each tensor mode corresponds to a dimension to be selected, the search space is the Cartesian product of candidate dimensions across modes, and exhaustive search is costly. Common strategies include iterative per-mode selection (fixing other mode dimensions and optimizing the current mode) or setting thresholds based on the eigenvalue spectrum of the reduced covariance matrix. In asymptotic theory, consistency of dimension selection requires that the candidate dimension set be bounded and the true dimensions identifiable, while the penalty term must grow appropriately with sample size. Improper dimension selection leads to underfitting (losing structural information) or overfitting (introducing noise directions), making this step crucial to PMTR's practical performance.

Potential Applications and Limitations

PMTR is suitable for scenarios where the response distribution is asymmetric, heavy-tailed, or contains outliers, such as financial risk prediction and medical imaging analysis. However, the paper currently focuses mainly on theoretical properties, and details of numerical experiments and real-world application cases are not fully elaborated in the abstract. Moreover, the computational complexity of modal regression may be higher than that of mean regression, and its efficiency on large-scale tensors remains to be further verified.

Overall, PMTR provides a robust and structure-preserving new tool for high-dimensional tensor regression and merits attention in future empirical research.

One-line Summary

Partial Modal Tensor Regression uses conditional modes to construct a tensor covariance that guides dimension reduction, achieving robust and structure-preserving high-dimensional tensor regression.

Paragraph Summary

This article introduces a new high-dimensional tensor regression method—Partial Modal Tensor Regression (PMTR). Traditional tensor regression mostly focuses on conditional means or conditional quantiles and performs poorly when the response distribution is asymmetric, heavy-tailed, or contains outliers. PMTR instead targets the conditional mode (the peak of the probability density) as the regression objective, implemented in two steps: first constructing a partial modal tensor covariance to select the tensor subspace most relevant to the response mode, then building a modal regression model in the reduced tensor space, thereby avoiding the curse of dimensionality caused by vectorization and preserving the intrinsic tensor structure. The paper proves consistency and asymptotic properties of the estimator under regularity conditions and provides a data-driven dimension selection procedure. The method is applicable to scenarios such as financial risk prediction and medical imaging analysis, though its computational efficiency and empirical performance remain to be verified.