CartoBoost NeuralEmbeddingFeatures

Use NeuralEmbeddingFeatures when learned ID vectors should become columns for another estimator. This page is for feature generation and ablation studies, not for serving a neural model directly.

When To Use

• Stable high-cardinality IDs recur across train and prediction rows.
• You want to compare base features against base features plus ID embeddings.
• Several downstream models should use the same train-side embedding columns.
• You can report repeated-ID and cold-ID splits separately.

Python Example

from cartoboost.neural import NeuralEmbeddingFeatures

features = NeuralEmbeddingFeatures(dim=8, random_state=7)
features.fit(train_ids, train_target)

train_embedding_cols = features.transform(train_ids)
valid_embedding_cols = features.transform(valid_ids)

Append the generated columns to the dense features used by CartoBoostRegressor, LightGBM, XGBoost, or another baseline. Fit the embedding table only on training rows when evaluating a holdout.

Interactive Example

Neural embedding feature browser model

Runs embedding in the browser with the bundled CartoBoost Wasm model.

Ready to run in this page.

Use When

Need	Better surface
Serve the supervised ID model directly.	NeuralEmbeddingStandaloneRegressor
Generate reusable embedding columns.	NeuralEmbeddingFeatures
Fit embeddings and a CartoBoost regressor together.	NeuralEmbeddingRegressor

Validation

If validation rows contain IDs observed during training, describe the result as repeated-ID generalization. If validation rows contain unseen IDs, report the fallback behavior explicitly.

Limitations

• Generated columns inherit the embedding model's cold-ID behavior.
• Fit the transformer on training rows only to avoid target leakage.
• Downstream feature importance does not explain the internal embedding dimensions semantically.