Implementation:Recommenders team Recommenders EmbDotBias Training

Knowledge Sources	Recommenders
Domains	Collaborative Filtering, Model Training, PyTorch
Last Updated	2026-02-10 00:00 GMT

Overview

The Trainer class and predict_rating function provide the training loop and single-item prediction utilities for the EmbeddingDotBias collaborative filtering model.

Description

The Trainer class wraps a PyTorch model with an AdamW optimizer (betas=(0.9, 0.99), eps=1e-5) and MSELoss loss function. It handles automatic device placement to GPU when available. The train_epoch method runs a single epoch of forward/backward passes over batched data, computing MSE loss between predicted and actual ratings. The validate method evaluates the model on a validation set using torch.no_grad() for memory efficiency, returning None if the validation set is empty. The fit method orchestrates multi-epoch training with per-epoch logging of both training and validation losses. The standalone predict_rating function takes a trained model along with a user ID and item ID, converts them to embedding indices via the model's _get_idx method, and returns a single predicted rating score, with error handling that logs failures and returns None.

Usage

Use the Trainer class when training an EmbeddingDotBias model on user-item rating data. It provides a complete training pipeline that handles optimizer configuration, loss computation, and device management. Use predict_rating for generating individual rating predictions after training, such as in an API endpoint or interactive evaluation scenario.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/embdotbias/training_utils.py
• Lines: 1-138

Signature

class Trainer:
    def __init__(self, model, learning_rate=1e-3, weight_decay=0.01)
    def train_epoch(self, train_dl)
    def validate(self, valid_dl)
    def fit(self, train_dl, valid_dl, n_epochs)

def predict_rating(model, user_id, item_id)

Import

from recommenders.models.embdotbias.training_utils import Trainer, predict_rating

I/O Contract

Inputs

Name	Type	Required	Description
model	torch.nn.Module	Yes	The PyTorch model to train (typically EmbeddingDotBias)
learning_rate	float	No	Learning rate for AdamW optimizer (default 1e-3)
weight_decay	float	No	Weight decay regularization for AdamW optimizer (default 0.01)
train_dl	DataLoader	Yes	Training data loader yielding (users_items, ratings) batches
valid_dl	DataLoader	Yes	Validation data loader yielding (users_items, ratings) batches
n_epochs	int	Yes	Number of training epochs
user_id (predict_rating)	str	Yes	The ID of the user for prediction
item_id (predict_rating)	str	Yes	The ID of the item for prediction

Outputs

Name	Type	Description
train_epoch return	float	Average training loss for the epoch
validate return	float or None	Average validation loss, or None if validation set is empty
predict_rating return	float or None	Predicted rating score, or None if an error occurs

Usage Examples

Basic Usage

from recommenders.models.embdotbias.model import EmbeddingDotBias
from recommenders.models.embdotbias.training_utils import Trainer, predict_rating

# Build the model
classes = {"userID": user_ids, "itemID": item_ids}
model = EmbeddingDotBias.from_classes(n_factors=40, classes=classes, y_range=(1, 5))

# Create trainer and fit
trainer = Trainer(model, learning_rate=1e-3, weight_decay=0.01)
trainer.fit(train_dl, valid_dl, n_epochs=5)

# Predict a single rating
rating = predict_rating(model, user_id="user_42", item_id="item_101")