Partial Modal Tensor Regression

Jintao Han, Li Zhang, Xuejun MaarXiv 2026 · arXiv

TL;DR

Partial Modal Tensor Regression (PMTR) is a new statistical framework designed to handle high-dimensional, multi-way data (tensors) when the relationship you're modeling is messy—specifically when it’s skewed, heavy-tailed, or full of outliers. By focusing on the conditional mode (the most probable outcome) rather than the average, and using a structure-preserving dimension reduction technique, PMTR provides a robust and interpretable way to analyze complex datasets like air quality or medical imaging.

What Problem Does This Paper Address?

In modern data science, we often deal with tensors—multi-dimensional arrays like a series of hourly pollutant readings across multiple weather variables. The standard way to handle these is "vectorization" (flattening the array into a long list), but this is a disaster for two reasons:

  1. The Curse of Dimensionality: The number of parameters explodes, leading to overfitting.
  2. Loss of Structure: You lose the inherent relationships between "hours of the day" and "types of pollutants."

Furthermore, most regression models look for the conditional mean (the average). However, if your data is skewed (like income or pollution levels), the average is easily pulled away by outliers. This paper addresses the need for a regression method that respects the multi-way structure of tensor data while being robust enough to target the mode—the "typical" or most likely value—rather than just the average.

Methods

The authors introduce Partial Modal Tensor Regression (PMTR). The workflow consists of three main stages:

  1. Partial Modal Tensor Covariance: The authors define a new way to measure the relationship between a tensor predictor and a response variable based on a "modal score." This score essentially weights observations based on how close they are to the most probable value.
  2. Dimension Reduction (Tensor PLS): Instead of using all the data at once, they employ a Tucker decomposition guided by a Partial Least Squares (PLS) criterion. This extracts the most "response-relevant" low-dimensional slices of the tensor, significantly reducing the number of parameters without losing the multi-way geometry.
  3. Modal Regression in Reduced Space: Once the data is projected into a smaller "latent" space, they perform the final regression.
  4. Rank Selection: They use a Ridge-type Eigenvalue Ratio (RER) criterion to automatically determine how many dimensions to keep for each "mode" (e.g., how many temporal patterns vs. how many pollutant patterns).

Datasets

The paper validates PMTR using two primary sources:

  • Synthetic Simulations: The authors generated 50x50 tensors with specific geometric coefficient patterns (squares, T-shapes, crosses, circles, and triangles). They tested these against various noise types: Gaussian (standard), $t_3$ (heavy-tailed), Log-normal (skewed), and Contaminated Normal (outliers).
  • Beijing Multi-Site Air Quality Data: A real-world dataset from the UCI Machine Learning Repository. It includes hourly measurements of pollutants ($SO_2, NO_2, CO, O_3$) and meteorological variables (Temp, Pressure, Dew Point, Wind Speed) from the Wanliu monitoring station (2013–2017).

Key Insights and Ideas

  • Mode > Mean for Skewed Data: In the Beijing air quality analysis, the distribution of $PM_{2.5}$ is heavily right-skewed. The "average" pollution level is often misleading because a few extreme days pull it upward. PMTR targets the "most likely" pollution level, which is more representative of a typical day.
  • Structure Matters: By treating the predictors as a 24 (hours) x 8 (variables) matrix rather than a 192-length vector, the model identifies specific "hotspots" in time. For example, $NO_2$ has a strong positive association with the $PM_{2.5}$ mode around 8:00 PM, a detail that might be blurred in a flattened model.
  • Robustness to "Noise": The simulation showed that PMTR remains stable even when the noise level is 200% of the signal. Because it uses a kernel-based modal objective function, it effectively ignores outliers that would ruin a standard Least Squares model.

Key Results

  • Consistency: The authors provide mathematical proofs showing that their estimators converge to the true values at an optimal rate as the sample size increases.
  • Pattern Recovery: In simulations, PMTR successfully recovered complex shapes (like a rank-10 bi-circle) even with heavy-tailed error distributions where traditional methods often fail.
  • Air Quality Findings: When applied to the Beijing data, the model reduced the 24-hour temporal dimension to just 8 effective dimensions. It revealed that $NO_2, O_3$, Temperature, and Pressure are the most stable predictors of the $PM_{2.5}$ mode, with their influence shifting significantly throughout the day (e.g., Temperature has a negative effect at 9 AM but a positive one at 2 PM).

Limitations and Future Work

  • Independence Assumption: The current model assumes observations are independent. In reality, air quality data has "serial dependence" (today's air depends on yesterday's). The authors had to sub-sample the data to seven-day blocks to mitigate this.
  • Sparsity: While the model reduces dimensions, it doesn't necessarily zero out irrelevant individual variables within those dimensions. The authors suggest that adding sparse regularization (like Lasso) could make the results even more interpretable.
  • Uncertainty Quantification: The paper focuses on point estimates. Future work is needed to develop formal hypothesis testing and confidence intervals for the tensor coefficients.

Summary updated

Abstract

With the rapid development of modern scientific research, tensor data have become increasingly prevalent across a wide range of fields. How to effectively characterize the relationship between tensor predictors and a response while preserving the intrinsic structure of tensor data has therefore become an important problem in high-dimensional statistical analysis. Existing tensor regression methods mainly focus on conditional means or conditional quantiles. When the response distribution is asymmetric, heavy-tailed, or contaminated by outlying observations, the conditional modal provides a robust and interpretable alternative characterization. In this paper, we propose a partial modal tensor regression (PMTR) method. The proposed method constructs the partial modal tensor covariance, which incorporates information on the conditional modal to guide tensor dimension reduction, and subsequently establishes a modal tensor regression model in the reduced tensor space. This strategy alleviates the dimensionality burden caused by direct vectorization while preserving the intrinsic structural information of tensor predictors. Under suitable regularity conditions, we establish the consistency and asymptotic properties of the proposed estimators and employ a data-driven dimension-selection procedure to determine the effective dimension along each tensor mode.
BibTeX
@misc{han2026partial,
  author = {Jintao Han and Li Zhang and Xuejun Ma},
  title = {{Partial Modal Tensor Regression}},
  year = {2026},
  eprint = {2610.34008},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34008},
}
Figure