Task-Conditioned Synthetic Data Generation for Improving Machine Learning Performance in Agricultural Prediction Tasks

Technical Summary: Task-Conditioned Synthetic Data Generation for Agricultural Prediction

Problem Statement
Machine Learning (ML) algorithms are increasingly used for agricultural prediction tasks, such as crop yield estimation and crop type classification. However, their performance is heavily constrained by the quantity and quality of available training data. Agricultural reference data often suffer from limitations including scarcity, measurement inaccuracies, incomplete records, and poor spatial-temporal coverage. While Synthetic Data Generation (SDG) offers a potential solution to augment datasets, existing SDG methods often fail to improve downstream ML performance. This is because many current approaches focus on mimicking the joint distribution of the original data (statistical fidelity) rather than ensuring the generated samples preserve the specific feature-target relationships required by the downstream task. Consequently, synthetic data may be statistically realistic but task-misaligned, leading to degraded model performance when used for training.

Methodology: TCSDG Algorithm
The paper proposes Task-Conditioned Synthetic Data Generation (TCSDG), a framework designed to generate synthetic tabular data that is explicitly optimized for downstream supervised learning tasks. Unlike unconditional generators, TCSDG integrates a "teacher–student" knowledge transfer paradigm and in-context learning. The algorithm operates through five distinct stages:

1. Budget Allocation: The total synthetic budget is distributed across the target space (classes for classification, bins for regression) using a hybrid prior. This combines the empirical class distribution with an inverse-frequency prior to ensure minority classes or underrepresented target regions receive proportional synthetic support without distorting the global marginal distribution.
2. Teacher Training and Calibration: A TabICL (Tabular Foundation Model for In-Context Learning) model is trained on the real data to act as a "task-aware critic." It does not generate data but evaluates candidates. For classification, it assesses assigned-class probability, prediction margin, and entropy. For regression, it evaluates the residual between the predicted mean and the candidate's target value. Thresholds for these metrics are calibrated using the real training data.
3. Candidate Generation: A large pool of candidate synthetic samples is generated using a Bayesian Network (BN). For classification, class-conditional BNs are fitted; for regression, a joint BN is fitted to features and targets. This pool is significantly oversized (at least 6x the required budget) to allow for aggressive filtering.
4. Teacher-Guided and Coverage-Preserving Selection: The final synthetic dataset is formed by the union of two subsets:
   • Teacher-Consistent Subset: Candidates that satisfy the task-specific criteria defined by the TabICL teacher (e.g., high confidence, low uncertainty).
   • Coverage Subset: Candidates selected from the remaining pool to ensure distributional support and diversity, using MiniBatch k-means clustering to select representatives closest to centroids.
5. Sample Weighting: Synthetic samples are assigned weights based on their teacher-derived quality scores. This allows downstream models (specifically those supporting weighted training like Random Forest and SVM) to prioritize high-confidence synthetic records while preserving the relative contribution of real data.

Experimental Setup
The study evaluated TCSDG against six established tabular SDG benchmarks (BN, CTGAN, ADSGAN, TVAE, RTVAE, and DDPM) across twelve supervised agricultural tasks:

• Six Crop Type Classification tasks using the LUCAS 2022 dataset across six European countries (Greece, Italy, Austria, Germany, Romania, Bulgaria).
• Six Crop Yield Prediction tasks (maize and wheat) using the CY-Bench dataset across six countries (Argentina, Germany, Zambia, India, France, Italy).

The evaluation considered two training regimes (full 70% data and reduced 30% data), four multiplication ratios (1x to 8x), and three downstream ML learners (Random Forest, MLP, SVM). Performance was measured using Balanced Accuracy (BAC) for classification and Root Mean Squared Error (RMSE) for regression.

Key Results

• Superior Performance: TCSDG was the only method to consistently improve ML performance across both tasks at the aggregate level. It improved performance in 89% of crop type classification experiments and 74% of crop yield prediction experiments when synthetic data was merged with real data.
• Synthetic-Only Capability: In the "synthetic-only" regime (where real training data was removed), TCSDG outperformed all other methods. It exceeded the baseline in 91% of classification comparisons, whereas other methods largely failed to recover task-relevant structure without real data.
• Robustness to Multiplication Ratio: Unlike benchmark methods, whose performance degraded as the synthetic-to-real ratio increased (due to dilution of real data by low-quality synthetic samples), TCSDG maintained stable or improved performance up to a ratio of 4x, after which it saturated rather than collapsed.
• Learner Dependency: The benefits of TCSDG were most pronounced for Random Forest and SVM, which could utilize the quality-aware sample weights. The Multi-Layer Perceptron (MLP) showed more modest gains, likely due to its inability to leverage these weights and higher sensitivity to imperfect calibration.
• Data Scarcity vs. Abundance: Contrary to the expectation that SDG is only useful in data-scarce settings, TCSDG maintained its advantage even when the real training data fraction increased from 30% to 70%, suggesting it provides complementary, non-redundant information.

Significance and Claims
The paper claims that the primary contribution of TCSDG is the shift from distributional fidelity to task utility. The authors argue that for agricultural prediction, synthetic data is only valuable if it preserves the conditional structure required by the downstream classifier or regressor.

The study demonstrates that:

1. Task-conditioning is critical: Simply generating realistic data is insufficient; the generation process must be guided by a task-aware critic to ensure the synthetic samples are informative for the specific prediction goal.
2. A unified pipeline is necessary: The integration of budget allocation, teacher-guided filtering, and sample weighting into a single pipeline is essential for success, as evidenced by the poor performance of the standalone Bayesian Network generator used within TCSDG when tested as a baseline.
3. Practical Framework: TCSDG offers a practical, extensible framework for precision agriculture that can be applied to diverse datasets and tasks, improving model robustness without requiring costly new data acquisition.

The authors conclude that carefully designed, task-conditioned synthetic data can effectively address data limitations in agriculture, provided the generation process prioritizes downstream utility over mere statistical mimicry.