Model Training

Model Training#

Basic Training#

All CoreRec models follow the same fit() API:

from corerec.engines.collaborative import SAR

model = SAR(similarity_type='jaccard')
model.fit(train_df)

For deep learning models:

from corerec.engines import DeepFM

model = DeepFM(
    embedding_dim=64,
    hidden_layers=[256, 128],
    epochs=20,
    learning_rate=0.001,
    batch_size=256,
    device='cuda',
)
model.fit(user_ids=user_ids, item_ids=item_ids, ratings=ratings)

Choosing the task: implicit ranking vs. rating prediction#

The deep CTR/graph models (DCN, DeepFM, GNNRec) accept a task argument that controls how labels are interpreted and how the model is trained:

from corerec.engines import DCN

# top-K recommendation (default): observed interactions are positives and the
# model is trained with negative sampling -> use this for ranking.
model = DCN(task="auto", num_negatives=4)        # 'auto' == 'implicit'

# rating prediction: regress the supplied rating with a linear head + MSE.
model = DCN(task="rating")

Important

Train these models with task="implicit" (the default) for recommendation. Passing raw 1–5 ratings to an implicit model without negatives is the classic mistake that makes the output collapse to a constant; CoreRec now handles this for you via negative sampling and warns if a trained model’s scores have (near-)zero variance. Use task="rating" only when you genuinely want to predict the rating value.

Training strong models for large-scale ranking#

For million-scale implicit-feedback ranking, CoreRec ships native, GPU, sparse trainers that produce strong embeddings and feed straight into the online serving engine:

from corerec.serving import OnlineRecommender

# trains a sparse LightGCN (SOTA-class on graph benchmarks) and builds the index
rec = OnlineRecommender.from_interactions(df, model="lightgcn", device="cuda")

See the Online Serving guide for the full train-and-serve path.

Callbacks#

CoreRec provides training callbacks for monitoring and control:

from corerec.utils.training_utils import EarlyStopping, ModelCheckpoint

# Stop training when validation loss plateaus
early_stop = EarlyStopping(patience=5, min_delta=0.001)

# Save the best model during training
checkpoint = ModelCheckpoint(
    filepath='artifacts/best_model',
    monitor='val_loss',
    save_best_only=True,
)

Validation#

Use the built-in validation helpers before training:

from corerec.utils.validation import validate_fit_inputs

# Validates data format, column names, and types
validate_fit_inputs(train_df, col_user='userID', col_item='itemID')

GPU Training#

Deep learning models support GPU acceleration:

model = DeepFM(device='auto')  # default: CUDA, then Apple MPS, then CPU
model = DeepFM(device='cuda')  # NVIDIA GPU
model = DeepFM(device='mps')   # Apple Silicon GPU
model = DeepFM(device='cpu')   # Force CPU
Contents