denograd 1.0.0

Instance noise reduction framework based on Deep Learning gradients agnostic to the network architecture.

DenoGrad: Deep Gradient Denoising Framework

📄 Description

DenoGrad is a novel gradient-based denoising framework designed to enhance the robustness and performance of Artificial Intelligence models, with a specific focus on interpretable (white-box) models.

Unlike conventional techniques that simply remove noisy instances or significantly alter the data distribution, DenoGrad leverages the gradients of a reference Deep Learning (DL) model—trained on the target data—to dynamically detect and correct noisy samples.

🚀 Key Features

• Gradient-Based Correction: Utilizes gradient information from deep models to guide the noise reduction process effectively.
• Distribution Preservation: Corrects instances while maintaining the original data distribution, avoiding oversimplification of the problem space.
• Task Agnostic: Validated effectively on both tabular data and time-series datasets.
• Interpretable AI Enhancement: Specifically engineered to boost the performance of interpretable models in noisy environments without sacrificing transparency.

🛠️ Installation

pip install denograd

📖 Basic Usage

import torch
import torch.nn as nn
from denograd import DenoGrad

# Initialize the denoiser with your reference model
denoiser = DenoGrad(
    model=my_deep_model, # DL model fitted to noisy data
    criterion=nn.MSELoss(),
)

# 1. Fit the noisy data
# For Tabular Data
denoiser.fit(
    X=x_noisy, 
    y=y_noisy, 
    is_ts=False
)

# For Time Series (requires window_size)
# denoiser.fit(
#    X=x_noisy,
#    y=['y'],
#    is_ts=True,
#    window_size=24,
#    future=1,
#    stride=1
# )

# 2. Denoise the dataset
x_clean, y_clean, x_gradients, y_gradients = denoiser.transform(
    nrr = 0.05, # Noise Reduction Rate. Same functionality as learning rate but for denoising porposes.
    nr_threshold = 0.01, # "Level" of noise allowed.
    max_epochs = 200, # Max number of epochs to perform the denoising process.
    denoise_y = True, # Enable denoising for the target variable (recommended for Tabular).
    batch_size = 1024,
    save_gradients = True # Save all the gradients calculated through the denoising process.
)

📝 Citation

If you use DenoGrad in your research, please cite our paper:

@article{denograd2025, title={DenoGrad: Deep Gradient Denoising Framework for Enhancing the Performance of Interpretable AI Models}, author={Alonso-Ramos, J. Javier and [Other Authors]}, journal={arXiv preprint arXiv:2511.10161}, year={2025} }