Implementation:Recommenders team Recommenders SARSingleNode Fit

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

Overview

Concrete tool for training a SAR collaborative filtering model on user-item interaction data provided by the recommenders library.

Description

The SARSingleNode.fit method is the main training entry point for the SAR model. Given a pandas DataFrame of user-item-rating interactions (with no duplicates), it performs the following steps:

1. Validates that the input contains no duplicate (user, item) pairs.
2. Generates continuous index mappings for users and items if not already set.
3. Validates that the rating column is numeric.
4. Optionally applies time-decay weighting to ratings based on timestamps.
5. Builds the user-affinity sparse matrix via compute_affinity_matrix.
6. Computes the item co-occurrence matrix via compute_cooccurrence_matrix.
7. Extracts item frequencies from the co-occurrence diagonal.
8. Derives the item-similarity matrix using the configured similarity metric (jaccard, cosine, lift, cooccurrence, inclusion index, mutual information, or lexicographers mutual information).
9. If normalization is enabled, computes an additional unity-rating affinity matrix for score scaling.

After fit completes, the model's user_affinity, item_similarity, and item_frequencies attributes are populated.

Usage

Call this method after instantiating a SARSingleNode model and before calling recommend_k_items or predict. The input DataFrame must contain user, item, and rating columns, and optionally a timestamp column if time-decay is enabled.

Code Reference

Source Location

• Repository: recommenders
• File: recommenders/models/sar/sar_singlenode.py
• Lines: L226-L323

Signature

def fit(self, df: pd.DataFrame) -> None

Import

from recommenders.models.sar.sar_singlenode import SARSingleNode

model = SARSingleNode(similarity_type="jaccard")
model.fit(train_df)

I/O Contract

Inputs

Name	Type	Required	Description
df	pd.DataFrame	Yes	User-item-rating interaction data. Must contain columns matching col_user, col_item, and col_rating (as configured in __init__). Must not contain duplicate (user, item) pairs. If timedecay_formula is True, must also contain a col_timestamp column.

Outputs

Name	Type	Description
return	None	The method modifies the model instance in-place. After fitting, the following attributes are populated:
self.user_affinity	scipy.sparse matrix	Sparse user-affinity matrix of shape (n_users, n_items).
self.item_similarity	scipy.sparse matrix	Sparse item-similarity matrix of shape (n_items, n_items).
self.item_frequencies	numpy.ndarray	Array of item frequency counts (diagonal of the co-occurrence matrix).

Usage Examples

Basic Usage

from recommenders.models.sar.sar_singlenode import SARSingleNode

# Initialize and train the model
model = SARSingleNode(
    col_user="userID",
    col_item="itemID",
    col_rating="rating",
    similarity_type="jaccard",
)
model.fit(train_df)

# After fit, model attributes are populated
print(model.user_affinity.shape)    # (n_users, n_items)
print(model.item_similarity.shape)  # (n_items, n_items)
print(model.item_frequencies.shape) # (n_items,)

# Training with time decay
model_td = SARSingleNode(
    col_user="userID",
    col_item="itemID",
    col_rating="rating",
    col_timestamp="timestamp",
    similarity_type="cosine",
    timedecay_formula=True,
    time_decay_coefficient=15,
)
model_td.fit(train_df)

Dependencies

• numpy - Numerical operations, dtype validation
• pandas - DataFrame input handling and manipulation
• scipy.sparse - Sparse matrix construction and multiplication

Related Pages

Implements Principle

• Principle:Recommenders_team_Recommenders_SAR_Model_Training

Requires Environment

• Environment:Recommenders_team_Recommenders_Python_Core_Dependencies