Implementation:Recommenders team Recommenders NRMSModel Init

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

Overview

Concrete tool for instantiating the NRMS (Neural News Recommendation with Multi-Head Self-Attention) model, including loading pre-trained word embeddings, building the news encoder, user encoder, and scorer.

Description

NRMSModel.__init__ initializes the full NRMS model through a two-phase process:

Phase 1 (NRMSModel-specific):

• Loads the pre-trained word embedding matrix from the numpy file specified in hparams.wordEmb_file via _init_embedding.
• Stores the embedding as self.word2vec_embedding.

Phase 2 (BaseModel.__init__):

• Sets the random seed for TensorFlow and NumPy.
• Creates train and test iterators from the provided iterator_creator (typically MINDIterator).
• Configures the TensorFlow session with GPU memory growth.
• Calls _build_graph, which in turn calls _build_nrms to construct:
  • A news encoder (word embedding -> dropout -> multi-head self-attention -> dropout -> additive attention).
  • A user encoder (TimeDistributed news encoder -> multi-head self-attention -> additive attention).
  • A training model that scores npratio + 1 candidates via dot product + softmax.
  • A scorer model that scores a single candidate via dot product + sigmoid.
• Compiles the training model with the configured loss function and optimizer.

Usage

Instantiate NRMSModel after preparing hyperparameters with prepare_hparams. The model is then ready for training via fit() and evaluation via run_eval().

Code Reference

Source Location

• Repository: recommenders-team/recommenders
• File (NRMSModel.__init__): recommenders/models/newsrec/models/nrms.py (lines 27-48)
• File (BaseModel.__init__): recommenders/models/newsrec/models/base_model.py (lines 29-77)

Signature

class NRMSModel(BaseModel):
    def __init__(
        self,
        hparams: HParams,
        iterator_creator: type,
        seed: int = None,
    ):
        """Initialization steps for NRMS.

        Args:
            hparams (HParams): Global hyper-parameters (head_num, head_dim, etc.).
            iterator_creator (type): Data loader class (e.g., MINDIterator).
            seed (int): Random seed for reproducibility.
        """

Import

from recommenders.models.newsrec.models.nrms import NRMSModel
from recommenders.models.newsrec.io.mind_iterator import MINDIterator

I/O Contract

Parameter	Type	Description
hparams	HParams	Validated hyperparameter object containing model architecture and training settings
iterator_creator	type	Data iterator class (e.g., MINDIterator) that will be instantiated for train/test
seed	int or None	Random seed for reproducible initialization

Attribute	Type	Description
self.model	keras.Model	Training model (multi-candidate scoring with softmax)
self.scorer	keras.Model	Inference model (single-candidate scoring with sigmoid)
self.newsencoder	keras.Model	News encoder submodel
self.userencoder	keras.Model	User encoder submodel
self.word2vec_embedding	numpy.ndarray	Pre-trained word embedding matrix
self.train_iterator	MINDIterator	Training data iterator (with negative sampling)
self.test_iterator	MINDIterator	Test/validation data iterator (no negative sampling)

Usage Examples

from recommenders.models.newsrec.models.nrms import NRMSModel
from recommenders.models.newsrec.io.mind_iterator import MINDIterator
from recommenders.models.newsrec.newsrec_utils import prepare_hparams

# Prepare hyperparameters
hparams = prepare_hparams(
    yaml_file,
    wordEmb_file=wordEmb_file,
    wordDict_file=wordDict_file,
    userDict_file=userDict_file,
    batch_size=32,
    epochs=5,
    head_num=20,
    head_dim=20,
    learning_rate=0.0001,
    npratio=4,
    title_size=30,
    his_size=50,
    data_format="news",
    support_quick_scoring=True,
)

# Initialize the NRMS model
model = NRMSModel(hparams, MINDIterator, seed=42)

# The model is now ready for training and evaluation
# model.fit(train_news, train_behaviors, valid_news, valid_behaviors)

Dependencies

• tensorflow — Deep learning framework (TF 1.x compatibility mode)
• tensorflow.keras — Keras API for building neural network layers and models
• numpy — For loading pre-trained word embeddings and numerical operations
• recommenders.models.newsrec.models.base_model.BaseModel — Parent class providing training, evaluation, and session management
• recommenders.models.newsrec.models.layers.AttLayer2 — Additive attention layer
• recommenders.models.newsrec.models.layers.SelfAttention — Multi-head self-attention layer

Related Pages

Implements Principle

• Principle:Recommenders_team_Recommenders_NRMS_Architecture

Requires Environment

• Environment:Recommenders_team_Recommenders_Python_Core_Dependencies
• Environment:Recommenders_team_Recommenders_GPU_CUDA_Environment

Uses Heuristic

• Heuristic:Recommenders_team_Recommenders_TensorFlow_Session_Ordering