Classification Estimators

Classification models map tensor features to discrete labels. They are optimized for stable decision boundaries and efficient inference.

Model structure

• Input tensors: [N × D] float32 arrays.
• Output tensors: class scores or label indices.
• Internally, classification reduces to margin-based or probabilistic objectives.

API usage

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

Internal behavior

• Logistic or softmax classifiers use fused native kernels where possible.
• Predictions are computed in a single pass over the feature tensor.
• Post-processing maps logits to label indices.

Performance implications

• Cross-entropy and softmax require stable accumulators.
• Use batch sizes that fit the available L2 cache.
• Avoid repeated label encoding in the hot path.

When to use

• Use classification estimators for categorical targets.
• Prefer them when inference latency matters.

When not to use

• Do not use classification estimators on regression targets.
• Do not use them when label imbalance mandates a specialized objective unless the estimator explicitly supports it.