Data Redundancy, Train-Test Leakage, and Cross-Source Generalisation in Machine-Learning Prediction of Fly Ash Concrete Compressive Strength

Machine-learning models for fly ash concrete are commonly developed from literature-assembled databases and evaluated using random train–test splits, which may overstate performance on previously unseen collection sources. This study examines how record redundancy, train–test contamination, and source heterogeneity affect reported performance in compressive-strength prediction. A database of 5,218 records from 14 collection sources was examined using record-level redundancy screening, operational domain screening, and 12 regression models. Of the combined operational total of 1,680 records (32.2%), 915 were reported as exact-value removals and are documented only at source-count level, whereas 765 near-duplicates were identified reproducibly from D1 at record level. In conventional random splits, 27.5% of test records had a near-duplicate counterpart in the training set. For these records, mean absolute error was 1.66 MPa for extreme gradient boosting (XGBoost) and 0.17 MPa for k-nearest neighbours, compared with 4.33 and 5.07 MPa, respectively, for uncontaminated records. After redundancy screening and source-eligibility filtering, 3,247 records from 11 collection sources remained. XGBoost achieved a macro-source coefficient of determination (R²) of 0.860 under source-stratified random splitting but only 0.369 under leave-one-source-cluster-out validation. Pairwise testing placed the five highest-ranked ensemble models within one Nemenyi group. The ten source-cluster blocks limit the strength of equivalence claims. The results indicate that data provenance, redundancy control, and source-aware validation affect apparent accuracy more strongly than model choice among the leading ensembles. Six minimum reporting requirements are proposed for machine-learning benchmarks assembled from concrete datasets.