🤖 machine learning

DenoGrad: A Gradient-Based Framework for Data Refinement in Tabular and Time-Series Learning

DenoGrad is a gradient-based framework that iteratively refines noisy tabular and time-series data by optimizing input values through a fixed pretrained neural network, improving downstream model performance without requiring clean reference data.

Original authors: J. Javier Alonso-Ramos, Ignacio Aguilera-Martos, Francisco Herrera, Andrés Herrera-Poyatos

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: J. Javier Alonso-Ramos, Ignacio Aguilera-Martos, Francisco Herrera, Andrés Herrera-Poyatos

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 trying to learn how to bake the perfect sourdough bread, but the recipe book you were given is a mess. Some of the measurements are smudged (noise in the features), and some of the instructions tell you to bake it for 50 hours instead of 50 minutes (noise in the targets).

If you just follow the messy book, your bread will be terrible. Usually, people try to fix this by throwing away the "bad" pages (filtering) or by guessing what the numbers should be based on math rules (statistical denoising).

This paper introduces DenoGrad, a new way to fix the recipe book using a "Master Chef" to guide the corrections.

The Core Idea: The "Master Chef" Method

Instead of using rigid math rules to guess the correct numbers, DenoGrad uses a Pre-trained Neural Network. Think of this network as a "Master Chef" who has already tasted thousands of loaves of bread. Even if the Chef was trained using some messy recipes, they still have a very strong "gut feeling" for what real bread should look and taste like.

Here is how DenoGrad works in three simple steps:

The Frozen Expert: We take our Master Chef (the neural network) and tell them: "Don't change your mind about what good bread is. Just stay exactly as you are." We "freeze" their knowledge.
The Error Check: We show the Chef a messy recipe. The Chef looks at it and says, "This recipe says bake for 50 hours. That's way off from what I know is true."
The Gradient Correction (The "Nudge"): Instead of the Chef trying to learn a new way of baking, we use the Chef's feedback to change the recipe itself. We use "gradients" (which you can think of as directional nudges) to push the smudged numbers and crazy instructions back toward the "correct" values that the Chef expects.

Why is this different?

It’s a "Reverse Attack": You might have heard of hackers using AI to "trick" a system by slightly changing data (an adversarial attack). DenoGrad is the opposite. It’s like a "Self-Correction Attack"—it uses those same mathematical nudges to push data away from errors and toward the truth.
It fixes both the "How" and the "What": Most methods only try to fix the ingredients (the features). DenoGrad fixes both the ingredients and the cooking time (the targets). It ensures the whole recipe makes sense together.
The "Consensus" for Time-Series: Imagine a recipe that says "Check the dough every 10 minutes." If one page says "Check at 10 mins" and another says "Check at 12 mins," it's confusing. For time-sensitive data (like stock markets or weather), DenoGrad uses a "Consensus Strategy." It looks at all the overlapping instructions and finds the "average agreement" before making a change, ensuring the timeline stays smooth and logical.

The Results: Better Data, Better Learning

The researchers tested this on everything from house prices to weather patterns and stock markets. They found that:

It’s a "Universal Polisher": It doesn't matter if you use a simple calculator or a super-advanced AI to bake the bread later; because the recipe is now cleaner, every baker performs better.
It preserves the "Soul" of the data: Some cleaning methods make data too "smooth," like turning a detailed photograph into a blurry smudge. DenoGrad is careful; it removes the "static" but keeps the important shapes and patterns intact.
It even helps "Clean" data: Surprisingly, even when the data wasn't very noisy, DenoGrad helped "tidy up" tiny inconsistencies, acting like a professional editor for a manuscript.

Summary in a Nutshell

DenoGrad is like having a highly experienced mentor look at your messy notes and, instead of teaching you new things, simply pointing at your mistakes and saying, "This number looks a bit too high; try nudging it down a bit." By doing this, the data becomes a much clearer map for any AI to follow.

Technical Summary: DENOGRAD

DENOGRAD is a novel, gradient-based framework designed for data refinement in tabular and time-series learning. It shifts the denoising paradigm from traditional statistical filtering or supervised reconstruction toward a model-guided, data-centric optimization approach.

1. The Problem: Limitations of Current Denoising

The authors identify three critical bottlenecks in existing data-denoising workflows:

Rigid Statistical Assumptions: Traditional methods (e.g., Kalman Filters, PCA, Wavelets) rely on fixed assumptions like Gaussianity or specific frequency distributions, which often fail in complex, real-world datasets.
Data Loss vs. Overcorrection: Filtering methods (removing noisy instances) reduce dataset diversity and size, while corrective methods (adjusting values) risk "overcorrecting" and distorting the original data distribution.
Supervision Requirements: Modern deep learning denoisers (e.g., Denoising Autoencoders) typically require paired "clean-noisy" datasets for training, which are rarely available in real-world scenarios.

2. Methodology: Gradient-Based Input Optimization

DenoGrad leverages the Spectral Bias of Deep Neural Networks (DNNs)—the tendency of networks to learn low-frequency, generalizable structures (the signal) before overfitting high-frequency stochastic variations (the noise).

The Core Mechanism:
Instead of updating model weights ( $\theta$ ), DenoGrad freezes a pre-trained predictive backbone and treats the input features ( $X$ ) and target variables ( $y$ ) as trainable parameters. It performs an "inverse adversarial optimization": while an adversarial attack maximizes loss to perturb inputs, DenoGrad minimizes loss to push corrupted observations toward the "clean manifold" learned by the model.

Key Technical Components:

Joint Feature-Target Optimization: Unlike methods that only refine features, DenoGrad optimizes $X$ and $y$ simultaneously. This prevents "gradient bias," where correcting features against a noisy, uncorrected target leads to suboptimal results.
Gating Mechanism (Tolerance Threshold $\tau$ ): To prevent over-smoothing and preserve valid stochasticity, a threshold $\tau$ is used. Updates are only applied to instances where the prediction error exceeds $\tau$ , ensuring high-confidence samples remain untouched.
Consensus Strategy (for Time-Series): To handle temporal dependencies in sequential data, DenoGrad uses a gradient accumulation buffer. Since a single time step $t$ appears in multiple overlapping sliding windows, the framework accumulates gradients from all windows and applies a single, temporally consistent average update.
Joint Normalization: Gradients for $X$ and $y$ are concatenated and normalized by their $L_2$ norm to ensure balanced updates across heterogeneous variables.

3. Key Contributions

Model-Independent Framework: It can be applied to any differentiable model (MLPs, CNNs, LSTMs) without requiring a secondary training phase or clean reference data.
Optimal Performance-Fidelity Trade-off: The framework achieves high predictive gains while maintaining the statistical topology of the original data (measured via Sliced Wasserstein Distance and Pearson correlation).
Universal Generalization: The refined data benefits a wide variety of downstream architectures, from linear models (Ridge) to gradient boosting (XGBoost) and specialized transformers (TabPFN).
Dataset-Level Regularization: DenoGrad acts as a regularizer even on "clean" data by harmonizing small inter-variable inconsistencies and suppressing latent noise.

4. Experimental Results

The authors tested DenoGrad on ten real-world datasets (5 tabular, 5 time-series) against seven state-of-the-art baselines (including PCA, Kalman Filters, and Denoising Autoencoders).

Predictive Gain: DenoGrad consistently outperformed baselines in Mean Squared Error (MSE) reduction. In time-series (e.g., ECL, Microsoft Stock), it achieved massive improvements (up to 98%).
Data Integrity: In Critical Difference (CD) diagrams, DenoGrad ranked highest for preserving data structure (lowest SWD) and feature correlation ( $\bar{\rho}$ ).
Ablation Success: The joint $X+y$ optimization was found to be significantly superior to $X$ -only optimization, particularly in time-series, where it provided a $3.4\times$ gain in predictive improvement.
Robustness: The framework remained effective across varying noise intensities ( $\sigma$ ) and showed stability across different backbone architectures.

5. Significance and Implications

DenoGrad represents a practical implementation of the Data-Centric AI paradigm. It demonstrates that predictive models should not merely be the "end product" of a pipeline but can serve as active tools for data curation. By using the implicit semantic knowledge encoded in a neural network to refine the very data it was trained on, DenoGrad bridges the gap between model-centric learning and data-centric engineering, providing a scalable way to improve dataset quality in production environments.