📊 statistics

Dynamic Frechet Regression with Feature Selection for Distributional Data

This paper proposes Dynamic Fréchet Regression (DFR), a novel framework that models index-dependent trajectories of distribution-valued responses by combining index-aware weighted Fréchet means with geometry-aware sparse metric learning to achieve accurate, interpretable predictions in high-dimensional settings.

Original authors: Kiran Adhikari, Amrutha Dinesh, Mathew Kuttolamadom, Ying Lin

Published 2026-07-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kiran Adhikari, Amrutha Dinesh, Mathew Kuttolamadom, Ying Lin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a chef trying to predict how a giant, magical cake will rise and change shape as it bakes. Usually, chefs just look at the final height or the average temperature. But what if the cake doesn't just rise; what if its entire personality changes? Maybe at the bottom layer, it's fluffy and wide, but by the top layer, it's dense and narrow, and the way it wobbles changes every second.

This is the problem scientists faced with a new kind of data called distributional data. Instead of a single number (like "5 inches tall"), the data is a whole cloud of possibilities—a full map of how something might behave, changing over time or depth.

The Old Way vs. The New Way

For a long time, scientists tried to study these changing clouds by squashing them into simple lines or averages. The paper argues that this is like trying to describe a symphony by only listening to the loudest instrument. You lose the harmony, the rhythm, and the surprise.

The authors explicitly argue against two common approaches:

Treating the data like a simple list of numbers: They say if you turn these complex clouds into a flat list of numbers, you break the natural shape of the data, like trying to fold a fluffy cloud into a square box.
Ignoring the "time" or "layer" aspect: Old methods treated every moment in the baking process as if it happened in a vacuum, ignoring that the cake at layer 5 is deeply connected to what happened at layer 4.

The Magic Solution: Dynamic Fréchet Regression (DFR)

The paper introduces a new tool called Dynamic Fréchet Regression (DFR). Think of DFR as a super-smart, time-traveling taste-tester.

Instead of guessing the cake's shape at a specific layer, DFR looks at all the cakes baked before. It asks: "Which past cakes look most like the one I'm trying to predict right now?"

But here is the clever part: DFR doesn't just look at the ingredients (the predictors). It also looks at when the cake was baked.

If you are predicting layer 10, DFR knows to pay extra attention to the cakes at layers 9 and 11, because they are neighbors.
It uses a special "distance ruler" (called the Wasserstein distance) that measures how different two clouds of data are without breaking their shape.

The paper suggests that by borrowing strength from these neighboring layers, DFR can predict the future shape of the cake much more accurately than the old methods, even when the data is noisy or messy.

Finding the Real Ingredients (Feature Selection)

In the kitchen, you might have 20 different spices, but only 6 actually change how the cake rises. The rest are just noise. The paper introduces a "magic sieve" (sparse metric learning) that automatically figures out which spices matter.

Instead of looking for a simple "amount" of spice (which doesn't work for these complex clouds), the method learns a special map of how the ingredients interact. If a spice doesn't change the map, the sieve removes it.

The Result: In their computer simulations, this sieve was incredibly good at finding the right ingredients. It never missed a true spice (100% recall), though it occasionally picked up a harmless extra one or two. This suggests the method is very safe: it's better to have a tiny bit of extra noise than to miss a crucial ingredient.

The Real-World Test: The Melting Metal Cake

To prove this wasn't just a computer game, the authors tested DFR on a real-world manufacturing process called Directed Energy Deposition (DED). Imagine a robot building a metal object layer by layer, melting metal with a laser.

The Predictors: The robot's settings: Laser Power, Scan Speed, and Dwell Time (how long it pauses).
The Response: The "melt pool"—the puddle of hot metal. Instead of just measuring the average temperature, they looked at the entire distribution of temperatures and the area of the puddle for every single layer (16 layers total) across 25 different builds.

The paper found that DFR was the best at predicting how these metal puddles would behave.

The Numbers: When predicting the "Peak Melt-pool Temperature," DFR made errors with an average of 0.037 (with a standard deviation of 0.016). This was lower (better) than the other methods, which had errors around 0.044 and 0.047.
The Discovery: The method correctly identified that the square of the Laser Power (how the power changes non-linearly) was the most important factor. It also found that for the "Melt-pool Area," the interaction between Laser Power and Scan Speed mattered.

What This Means

The paper doesn't claim to have solved every problem in the universe. It suggests that for data that evolves over time or depth—like manufacturing, weather patterns, or medical monitoring—this new "time-aware, shape-preserving" way of looking at data is a powerful upgrade.

It shows that by respecting the natural shape of the data and understanding how layers connect, we can predict complex, changing systems with greater accuracy and clarity than ever before. The authors are confident in their simulations and their real-world test, showing that this approach works better than the standard tools currently used in the field.

Technical Summary: Dynamic Fréchet Regression with Feature Selection for Distributional Data

Problem Statement
Contemporary scientific and engineering applications increasingly generate responses that are not scalar or vector values, but statistical objects—specifically probability distributions—whose forms evolve over an ordered index such as time, depth, or spatial layer. These "dynamic distributional trajectories" pose significant challenges for regression analysis. Standard Euclidean regression models are ill-suited because they discard intrinsic variability or require linearizing transformations that distort the natural geometry of probability spaces. While existing distributional regression methods, such as Global Fréchet Regression (GFR), respect the metric geometry of the response space (e.g., Wasserstein space), they are inherently static; they treat observations at different indices as exchangeable and fail to exploit cross-index dependence or smooth evolution. Furthermore, existing approaches lack principled mechanisms for feature selection in high-dimensional settings where explicit regression coefficients are unavailable for distribution-valued responses.

Methodology
The authors propose Dynamic Fréchet Regression (DFR), a framework designed to model the index-dependent evolution of distribution-valued responses driven by scalar predictors. The methodology consists of three core components:

Index-Aware Weighting Mechanism:
DFR extends GFR by introducing a weighting function that depends jointly on predictor similarity and index proximity. Unlike GFR, which uses weights constant across indices, DFR defines the prediction at a specific index $j$ as a weighted Fréchet mean in the metric space of distributions. The weights are constructed based on a roughness-penalized function-on-scalar regression analogy, allowing the model to borrow strength from neighboring indices while maintaining local specificity. A roughness parameter ( $\lambda$ ) controls the degree of cross-index smoothing: small values favor adaptive, index-specific predictions, while large values promote smoother trajectories.
Geometry-Aware Feature Selection:
To address the challenge of variable selection without explicit coefficients, DFR incorporates a sparse metric learning approach. Instead of shrinking coefficients, the method learns a sparse similarity metric matrix ( $M$ ) that governs how scalar inputs influence the similarity weights defining the Fréchet mean. By applying a group-lasso penalty to the rows of $M$ , the algorithm shrinks entire rows (and columns) to zero, effectively identifying and excluding nuisance predictors. This process is performed in a two-stage procedure: first, a sparse metric is learned to identify the active predictor set; second, the DFR model is refitted using only the selected predictors with the roughness parameter tuned via cross-validation.
Efficient Implementation for Univariate Distributions:
For univariate responses, the framework admits an efficient implementation in quantile space. Under the 2-Wasserstein metric, the estimation problem reduces to a weighted isotonic regression, which can be solved using the Pool Adjacent Violators Algorithm (PAVA). This ensures valid distributional outputs and scales efficiently to practical data sizes.

Key Contributions

Dynamic Fréchet Regression (DFR): A novel regression framework that models index-dependent trajectories of distribution-valued responses while preserving the intrinsic geometry of the response space (e.g., Wasserstein space).
Index-Aware Weighting: A theoretical extension of Global Fréchet Regression that incorporates index proximity into the weighting scheme, enabling the modeling of smooth evolution and cross-index dependence without imposing Euclidean linear structures.
Sparse Metric Learning for Feature Selection: A geometry-aware feature selection mechanism tailored for distributional data that identifies predictors driving distributional dynamics by learning a sparse similarity metric, avoiding the need for explicit regression coefficients.
Two-Stage Estimation Procedure: A stable workflow combining feature selection via sparse metric learning with subsequent refitting and roughness smoothing to handle high-dimensional and correlated predictor spaces.

Results
The paper evaluates DFR through simulation studies and an application to additive manufacturing data:

Simulation Studies:
- Feature Selection: DFR consistently achieved perfect recall (identifying all truly active predictors) across linear, quadratic, and periodic evolution scenarios. While precision varied slightly due to conservative selection strategies, the method effectively filtered out nuisance variables.
- Prediction Accuracy: DFR variants (with and without feature selection) significantly outperformed Global Fréchet Regression (GFR) and Function-on-Scalar Regression (FoSR) in terms of Mean Squared Prediction Error (MSE). The inclusion of index-aware weighting allowed DFR to capture dynamic evolution more accurately than static GFR. Feature selection further improved accuracy, particularly in scenarios with nonlinear and heterogeneous signal structures. FoSR, which treats distributions as Euclidean functional data, exhibited substantially higher errors, highlighting the necessity of preserving distributional geometry.
Additive Manufacturing Case Study:
- Applied to Directed Energy Deposition (DED) data involving 25 builds and 16 layers, DFR modeled the evolution of melt-pool temperature and area distributions.
- DFR-FS-RS (the full model with feature selection and smoothing) achieved the lowest test error and reduced variability compared to GFR and FoSR.
- The feature selection results were interpretable and physically consistent: the quadratic term of laser power was identified as the dominant driver for peak temperature, while melt-pool area depended on interactions between laser power, scanning speed, and dwell time.

Significance and Claims
The paper claims that DFR addresses a clear methodological gap by providing a framework that simultaneously respects the geometric structure of distributional responses and models their dynamic evolution across indices. The authors assert that existing methods either ignore dynamic structure (GFR) or rely on incompatible Euclidean assumptions (FoSR). By integrating index-aware weighting and sparse metric learning, DFR offers a stable, interpretable, and geometry-preserving approach for regression with dynamic distribution-valued data. The methodology is presented as applicable beyond additive manufacturing to any domain involving indexed distributional data, such as environmental monitoring, neuroscience, and reliability engineering, where quantifying and predicting variability is essential. The work emphasizes that modeling full distributions, rather than summary statistics, provides deeper insights into process behavior and uncertainty.