The problematic case of data leakage: A case for leave-profile-out cross-validation in 3-dimensional digital soil mapping
Bibliographic record
Abstract
• Data leakage in machine learning undermines model reliability. • Vertical autocorrelation in test datasets greatly compromise accuracy metrics. • Leave-profile-out cross-validation is needed to avoid data leakage in 3D models. • Validation methods for 3D models must be clearly reported in the literature. Data leakage occurs when there is an overlap between the data used for model fitting and hyperparameter tuning, and those used for testing. This overlap biases the model performance, making it uninformative regarding the model’s ability to generalize. This is a significant issue in machine learning and predictive soil mapping, compromising model reliability. To demonstrate this issue, the 3-dimensional (3D) digital soil mapping (DSM) approach, whereby depth is used as a predictor of soil properties, was investigated. We compare two common approaches from the literature: leave-sample-out cross-validation (LSOCV) versus leave-profile-out cross-validation (LPOCV). Here, we argue that LSOCV results in contamination of the test dataset due to the potential vertical autocorrelation of soil properties from different samples within the same profile, and a more appropriate approach for testing 3D DSM models should be to fully partition all soil samples from the same profile to either the training or test dataset (i.e., LPOCV). Using the Ottawa region of Ontario, Canada, as a case study, cation exchange capacity (CEC), clay content, pH, and total organic carbon (TOC) were predicted using machine learning, and the discrepancy in accuracy metrics was reported. Furthermore, we evaluated the effects of data augmentation (i.e., the creation of additional synthetic data points from the original data) on accuracy metrics, a common practice in 3D DSM. Here, it was shown that with the augmented dataset, LSOCV generated overly optimistic accuracy metrics (e.g., CCC) that were 29–62% higher than LPOCV, while for the non-augmented data, the accuracy metrics were 8–18% higher, suggesting that vertical autocorrelation had a strong influence on inflating model accuracy through data leakage. As such, we strongly urge DSM practitioners to provide greater clarity when describing how model accuracy metrics were ascertained and to consider the use of LPOCV when applied to 3D DSM. This brings about broader concerns that policymakers and stakeholders may use map products with the false impression that the maps are more accurate than they are. Future research should focus on refining DSM methods and considering data structure to prevent data leakage in modelling soil properties.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.184 | 0.291 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.002 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.003 | 0.005 |
| Scholarly communication | 0.003 | 0.004 |
| Open science | 0.004 | 0.004 |
| Research integrity | 0.004 | 0.005 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".