Principle:Recommenders team Recommenders News Model Training

Knowledge Sources	Recommenders
Domains	News Recommendation, Model Training, Deep Learning
Last Updated	2026-02-10 00:00 GMT

Overview

Training neural news recommendation models follows an epoch-based loop with per-epoch validation, where each training step computes impression-level loss over batches constructed with negative sampling (controlled by the npratio hyperparameter).

Description

The training principle for neural news recommenders encompasses the full training loop lifecycle:

Epoch-Based Training:

1. For each epoch (1 to hparams.epochs):
   1. Load training data from the news and behaviors files via the train iterator.
   2. The train iterator applies negative sampling with ratio npratio: for each positive (clicked) candidate, npratio negative (non-clicked) candidates are sampled from the same impression.
   3. For each mini-batch:
      • Extract input features (clicked title batch and candidate title batch) and labels.
      • Run one gradient update step using model.train_on_batch.
      • Accumulate the batch loss into the epoch loss.
      • Periodically log the running loss (every show_step steps).
   4. After all batches, compute the epoch training time.

Per-Epoch Validation:

1. At the end of each epoch, run run_eval on the validation set.
2. Log the validation metrics (AUC, MRR, NDCG@5, NDCG@10).
3. Optionally, if test files are provided, also evaluate on the test set.

Loss Function:

• For NRMS, the loss is categorical cross-entropy over the softmax of npratio + 1 candidate scores, where the positive candidate is the target class.
• This is equivalent to a listwise ranking loss that encourages the model to assign the highest probability to the clicked article.

Training Control:

• An optional step_limit parameter allows early termination within an epoch (useful for debugging or quick experiments).
• Progress is displayed via a tqdm progress bar showing step count, total loss, and data loss.

Usage

Use model training after dataset preparation, hyperparameter configuration, and model initialization. The fit() method is the standard entry point that handles the full training and validation loop. Training produces a model ready for final evaluation or prediction export.

Theoretical Basis

Negative Sampling

For each impression in training data:
  positive_candidate = clicked article
  negative_candidates = sample(npratio articles from non-clicked articles)

  If fewer non-clicked articles than npratio, pad with zeros.

  Input to model: [positive, neg_1, neg_2, ..., neg_npratio]
  Label: [1, 0, 0, ..., 0] (one-hot with positive at index 0)

Training Loss

scores = dot(candidate_news_vectors, user_vector)  # shape: (npratio + 1,)
probabilities = softmax(scores)
loss = categorical_cross_entropy(one_hot_label, probabilities)

The loss encourages:
  score(clicked_article) >> score(non_clicked_articles)

Training Loop Pseudocode

for epoch in range(1, epochs + 1):
    epoch_loss = 0
    step = 0
    for batch in train_iterator.load_data(train_news, train_behaviors):
        if step_limit and step >= step_limit:
            break
        loss = model.train_on_batch(batch_input, batch_label)
        epoch_loss += loss
        step += 1

    # Per-epoch validation
    eval_results = run_eval(valid_news, valid_behaviors)
    print(f"Epoch {epoch}: loss={epoch_loss/step:.4f}, AUC={eval_results['group_auc']:.4f}")

Related Pages

Implemented By

• Implementation:Recommenders_team_Recommenders_BaseModel_Fit