EPA Toxic Substance Emission Prediction: Building a Machine Learning Regulatory Pipeline Against Data Leakage

This article introduces a high-precision prediction system for the U.S. Environmental Protection Agency (EPA) Toxic Release Inventory (TRI) data, focusing on its two core innovations: systematically identifying and isolating 17 data leakage patterns, and using a two-level stacking ensemble learning strategy to improve prediction performance. The project aims to address data leakage issues in regulatory applications, ensure model generalization ability and credibility, and provide support for environmental regulatory decision-making, enterprise compliance management, and academic research.

In the field of environmental regulation, accurately predicting the toxic substance emissions of enterprises is crucial for policy formulation and compliance risk assessment. The EPA TRI dataset is a key public resource, but directly using it to train ML models carries significant data leakage risks (e.g., embedding components of the target variable into features), leading to inflated model performance and failure in real-world deployment. Building a robust pipeline that can identify and isolate leakage patterns is the core challenge in developing a reliable prediction system.

The project systematically identifies 17 potential data leakage patterns, including: target variable decomposition leakage (features contain components of the target), time-series look-ahead leakage (using future information to predict the present), inconsistent aggregation levels (mixing data of different granularities), derived feature leakage (using features calculated from the target as inputs), etc. Through a strict data audit process, a systematic method for detecting and eliminating leakage is established to ensure the authenticity of training data and the generalization ability of the model.

The project adopts an innovative two-level stacking method:

1. Level 1: Differential Evolution Optimized Weighted Mixing — The base layer uses heterogeneous models (e.g., gradient boosting trees, random forests, neural networks), and the prediction results are combined with optimal weights found via a differential evolution algorithm to maximize ensemble performance.
2. Level 2: Linear Regression Meta-Learning — The meta-learning layer takes the prediction results from Level 1 as input, learns the optimal combination method, preserves the diversity of base models, and reduces overfitting risks.

Data Processing Pipeline: The end-to-end process includes data acquisition (2022 EPA TRI data), cleaning and validation (quality checks, outlier/missing value handling), feature engineering (building features under anti-leakage constraints), model training (two-level stacking architecture and cross-validation), and evaluation monitoring (multi-metric assessment). Performance Metrics: On the log1p-transformed target variable, RMSE=0.2341 and R²=0.9966, indicating extremely high prediction accuracy and the ability to explain most of the variance in the target variable. The results are obtained under the premise of eliminating leakage, so they are highly credible.

Environmental Regulatory Decision Support: Early identification of enterprises/regions with abnormal emissions, optimization of regulatory resource allocation, and rapid screening of the rationality of enterprise self-reported data; Enterprise Compliance Management: Internal audit of data consistency, setting emission reduction targets based on industry benchmarks, and early warning of non-compliant operation links; Academic Research Value: The systematic method for handling data leakage provides a framework for similar fields, and the two-level stacking strategy demonstrates the potential of ensemble learning in structured data prediction.

The project uses modern data science tools: Marimo (interactive exploration and presentation), Conda (environment management), Python data science ecosystem (pandas, scikit-learn, etc.), and differential evolution optimization (possibly using scipy or dedicated libraries).

The epa-tri-ml project highlights the importance of handling data leakage in real-world data science. Its core value lies in establishing a reusable methodology: 1. Systematic thinking (treating leakage identification as a key engineering link); 2. Multi-layer defense (feature auditing + model architecture design to ensure data quality); 3. Balancing performance and credibility (high accuracy while ensuring interpretability and reliability). It has important reference value for developers of regulatory-level prediction systems.