Implementation:Recommenders team Recommenders Cornac BPR

Knowledge Sources	Recommenders
Domains	Recommendation Systems, Collaborative Filtering
Last Updated	2026-02-10 00:00 GMT

Overview

Extends the Cornac library's BPR (Bayesian Personalized Ranking) model with an efficient vectorized top-k recommendation method that returns results as a pandas DataFrame.

Description

The BPR class inherits from cornac.models.BPR and adds a recommend_k_items method for generating top-k item recommendations per user. Instead of scoring user-item pairs one at a time, the method computes all predictions simultaneously by multiplying user latent factor matrices (U) with item latent factor matrices (V) plus item biases (B) using the formula preds_matrix = U @ V.T + B. For efficient top-k selection, it uses np.argpartition to select the top-k items per user without fully sorting the entire prediction matrix. The method optionally removes previously seen user-item pairs from the training data by performing a left merge against the seen interactions.

Usage

Use this class when you need to generate top-k recommendations from a trained BPR model for evaluation with ranking metrics such as NDCG, MAP, or Precision@K. It bridges the Cornac BPR model with the recommenders library's pandas DataFrame-based evaluation pipeline.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/cornac/bpr.py
• Lines: 1-83

Signature

class BPR(CBPR):
    def __init__(self, *args, **kwargs)

    def recommend_k_items(
        self,
        data,
        top_k=None,
        remove_seen=False,
        col_user=DEFAULT_USER_COL,
        col_item=DEFAULT_ITEM_COL,
        col_prediction=DEFAULT_PREDICTION_COL,
    )

Import

from recommenders.models.cornac.bpr import BPR

I/O Contract

Inputs

Name	Type	Required	Description
data	pandas.DataFrame	Yes	DataFrame containing user and item columns from which to derive the recommendation scope
top_k	int	No	Number of items to recommend per user; defaults to all items if None
remove_seen	bool	No	If True, removes (user, item) pairs already seen in training data from results
col_user	str	No	Name of the user column; defaults to DEFAULT_USER_COL
col_item	str	No	Name of the item column; defaults to DEFAULT_ITEM_COL
col_prediction	str	No	Name of the prediction score column; defaults to DEFAULT_PREDICTION_COL

Outputs

Name	Type	Description
return	pandas.DataFrame	DataFrame with columns (col_user, col_item, col_prediction) containing top-k predicted items per user sorted by score

Usage Examples

Basic Usage

from recommenders.models.cornac.bpr import BPR
import cornac

# Train the BPR model
bpr_model = BPR(k=50, max_iter=100, learning_rate=0.001)
train_set = cornac.data.Dataset.from_uir(train_data.itertuples(index=False))
bpr_model.fit(train_set)

# Generate top-10 recommendations for all users
top_k_predictions = bpr_model.recommend_k_items(test_data, top_k=10, remove_seen=True)