Implementation:Recommenders team Recommenders DKN Item2Item Model

Knowledge Sources	Recommenders
Domains	Recommendation Systems, Deep Learning, Knowledge-Aware Recommendation
Last Updated	2026-02-10 00:00 GMT

Overview

Adapts the DKN (Deep Knowledge-aware Network) model for item-to-item recommendations, computing similarity between news articles using knowledge-aware document embeddings rather than user-to-article relevance.

Description

The DKNItem2Item class extends the base DKN model and overrides several key methods to support item-to-item recommendation.

The _build_dkn method generates L2-normalized document embeddings via _build_doc_embedding, reshapes them into groups of (source item, positive target, N negative targets), and computes dot-product similarity scores between the source item and all target items using tf.math.multiply followed by a sum reduction. The result is a tensor of shape (batch_size, neg_num + 1) stored as self.pred_logits.

The _build_doc_embedding method adds a dense tanh transformation layer on top of the KCNN (Knowledge-aware Convolutional Neural Network) output using a learned weight matrix W_doc_trans, producing dense document embeddings.

The _get_pred method applies softmax over the similarity scores. The _compute_data_loss method uses negative log-likelihood of the first (positive) item's softmax score as the training loss.

The eval method generates synthetic labels where the first item in each group is marked as positive (label=1), since ground truth is implied by item ordering. The run_eval method evaluates the model using pairwise metrics via cal_metric.

Usage

Use this model for finding similar news articles based on content and knowledge graph entity similarity. It is designed for item-to-item news recommendation scenarios as demonstrated in the KDD2020 tutorial.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/deeprec/models/dkn_item2item.py
• Lines: 1-129

Signature

class DKNItem2Item(DKN):
    def _compute_data_loss(self)

    def _build_dkn(self)

    def _get_pred(self, logit, task)

    def _build_doc_embedding(self, candidate_word_batch, candidate_entity_batch)

    def eval(self, sess, feed_dict)

    def run_eval(self, filename)

Import

from recommenders.models.deeprec.models.dkn_item2item import DKNItem2Item

I/O Contract

Inputs

Name	Type	Required	Description
sess (eval)	tf.Session	Yes	The TensorFlow model session object
feed_dict (eval)	dict	Yes	Feed values dictionary mapping graph elements to numpy arrays
filename (run_eval)	str	Yes	Path to evaluation data file containing news article IDs
candidate_word_batch (_build_doc_embedding)	tf.Tensor	Yes	Batch of word indices for candidate news articles
candidate_entity_batch (_build_doc_embedding)	tf.Tensor	Yes	Batch of entity indices for candidate news articles

Outputs

Name	Type	Description
return (_build_dkn)	tf.Tensor	Prediction logits of item-to-item similarity scores in shape (batch_size, neg_num + 1)
return (eval)	tuple(numpy.ndarray, numpy.ndarray)	A tuple of (predictions, labels) arrays for evaluation
return (run_eval)	dict	Dictionary containing pairwise evaluation metrics

Usage Examples

Basic Usage

from recommenders.models.deeprec.models.dkn_item2item import DKNItem2Item
from recommenders.models.deeprec.io.dkn_item2item_iterator import DKNItem2itemTextIterator
from recommenders.models.deeprec.deeprec_utils import prepare_hparams

# Prepare hyperparameters and create the model
hparams = prepare_hparams(yaml_file, news_feature_file=news_feature_file)
model = DKNItem2Item(hparams, DKNItem2itemTextIterator)

# Train the model
model.fit(train_file)

# Evaluate using pairwise metrics
eval_results = model.run_eval(test_file)
print(eval_results)  # e.g., {'auc': 0.85, 'ndcg': 0.72}