Regression Estimators

Regression models in PML compute continuous targets from numeric tensor inputs. They are designed for low-latency inference and efficient training.

Model structure

• Input features are represented as [N × D] float32 tensors.
• Targets are [N] float32 tensors.
• Regression training optimizes a differentiable loss or tree-based objective.

Data flow

Tensor samples [N × D]
   ├─ preprocessing / normalization
   └─ estimator.train() → loss / gradient computation

Example API usage

$dataset = Dataset::fromCSV('datasets/housing/train.csv', labelColumn: 0);
$model = new Pml\Estimators\Regression\GBDTRegressor();
$model->train($dataset);
$predictions = $model->predict($dataset);

Internal behavior

• Training may use native tensor operations for matrix algebra.
• Tree-based regressors may use C-backed histograms and split scoring.
• Weights and trees are persisted separately from PHP object structure.

Performance considerations

• Regression workloads benefit from dense tensors and contiguous layout.
• Use Dataset::materialize() before training to avoid repeated dataset conversion.
• Avoid broadcasting large tensors in the training loop unless required.

When to use

• Use regression estimators for real-valued prediction targets.
• Use tree-based regressors when feature normalization is less important.

When not to use

• Do not use regression estimators for categorical classification targets.
• Do not use dense regression models on extremely sparse high-dimensional data without prior dimensionality reduction.