Quantifying the impact of data quality on machine learning prediction: a large-scale study of construction product environmental declarations

Abstract Purpose Environmental product declaration (EPDs) are increasingly used for automated environmental assessments and machine learning (ML) applications. However, a systematic evaluation of how EPD data quality affects ML prediction reliability remains unexplored. Prior ML-EPD studies acknowledge data heterogeneity qualitatively but do not quantify its predictive impact, nor have they empirically established the cost-benefit tradeoff between data cleaning, harmonization and algorithmic sophistication, or the extent of cross-operator model generalizability. Method A dataset of 22,397 construction product EPDs from multiple program operators was systematically assessed along four data quality dimensions: completeness, temporal representatives, geographical representatives, and methodological consistency. Based on this assessment, a three scenario experimental design was implemented: (A) baseline predictions using EPDs with minimal cleaning (EPDs “as-is”); (B) rigorously cleaned and harmonized data, with standardizing schemas, units, product categories, life cycle modules, and EN 15804 versions, resulting in the exclusion of 28.66% of records on quality grounds; and (C) the same harmonized dataset combined with hyper parameter optimization and ML based external validation on held-out program operators. Five ML algorithms (Extreme Gradient Boosting (XGB), Random Forest (RF), Decision Tree (DT), Artificial Neural Network (ANN) and Support Vector Regression (SVR) were trained to predict Global Warming Potential (GWP). Results Data cleaning and harmonization significantly improved prediction accuracy (R² improvement from 0.30 to 0.88 and significant error reduction), whereas hyperparameter tuning on the harmonized dataset yielded only modest additional gains (ΔR² = 0.02). Completeness (18.8% data loss) and methodological consistency (7.3% loss) dominated data exclusions, directly affecting high-importance predictive features identified through SHAP analysis. Notably, SHAP explainability analysis revealed the program operator as a surprisingly influential predictor, acting as a proxy for latent cross-scheme methodological heterogeneity rather than true geographical variance. External validation on held-out program operators confirmed cross-scheme generalization (R² = 0.84), with feature rankings stable across training and external test sets. Conclusion This study provides the first empirical quantification of how EPD data quality and cross-operator harmonization thresholds (approximately 72% of records retained) govern the reliability of ML-based impact predictions. Upstream data quality investments are demonstrated to yield substantially greater returns than downstream algorithmic sophistication. Consequently, this work establishes that data curation and governance rather than merely compiling EPD repositories are prerequisite for reliable ML-based environmental assessments and provides a practical, reliable guideline for effectively cleaning and harmonizing heterogeneous EPD repositories.