Implementation:Recommenders team Recommenders GeoIMC Data

Knowledge Sources	Recommenders
Domains	Matrix Completion, Data Loading, Feature Engineering
Last Updated	2026-02-10 00:00 GMT

Overview

Provides data handling classes for the GeoIMC (Geometry-aware Inductive Matrix Completion) model, including filtered data access, base dataset preprocessing, and a MovieLens-100K specific implementation.

Description

This module contains three classes that form the data infrastructure for the GeoIMC model. DataPtr is a lightweight pointer class that holds a sparse CSR target matrix and entity feature arrays, supporting filtered access via configurable row and column indices through data_indices and entity_indices. Dataset is the base class that provides feature normalization (using either length normalization or sklearn's normalize for CSR matrices), PCA dimensionality reduction via the reduce_dims utility, and train/test splitting on row entities to enable cold-start evaluation scenarios. ML_100K extends Dataset to handle the MovieLens-100K dataset specifically: it reads user features (age, gender, occupation, zip code as one-hot encoded columns) and item features (movie title, release date, genres), converts ratings into COO sparse matrices with optional normalization or binarization target transforms, and loads pre-split train/test files.

Usage

Use this module when working with the GeoIMC inductive matrix completion model. The ML_100K class provides end-to-end data loading for MovieLens-100K, while the Dataset base class can be extended for other datasets. Use DataPtr when you need filtered views of the target matrix and entity features without copying data.

Code Reference

Source Location

• Repository: Recommenders
• File: recommenders/models/geoimc/geoimc_data.py
• Lines: 1-270

Signature

class DataPtr:
    def __init__(self, data, entities)
    def get_data(self)
    def get_entity(self, of="row")

class Dataset:
    def __init__(self, name, features_dim=0, normalize=False, target_transform="")
    def normalize(self)
    def generate_train_test_data(self, data, test_ratio=0.3)
    def reduce_dims(self)

class ML_100K(Dataset):
    def __init__(self, **kwargs)
    def df2coo(self, df)
    def load_data(self, path)
    def _load_user_features(self, path)
    def _load_item_features(self, path)

Import

from recommenders.models.geoimc.geoimc_data import DataPtr, Dataset, ML_100K

I/O Contract

Inputs

DataPtr.__init__

Name	Type	Required	Description
data	csr_matrix	Yes	The target data matrix in CSR sparse format
entities	Iterator	Yes	An iterator of 2 elements (ndarray) containing features for row and column entities

Dataset.__init__

Name	Type	Required	Description
name	str	Yes	Name of the dataset
features_dim	int	No	Dimension for PCA reduction; 0 means no reduction (default 0)
normalize	bool	No	Whether to normalize entity features (default False)
target_transform	str	No	Transform for target values: 'normalize', 'binarize', or for none (default )

Dataset.generate_train_test_data

Name	Type	Required	Description
data	csr_matrix	Yes	The entire target matrix
test_ratio	float	No	Ratio of test split (default 0.3)

ML_100K.load_data

Name	Type	Required	Description
path	str	Yes	Path to the directory containing the ML-100K dataset files

Outputs

DataPtr.get_data

Name	Type	Description
return	csr_matrix	Target matrix, optionally filtered by data_indices

DataPtr.get_entity

Name	Type	Description
return	numpy.ndarray	Entity feature matrix, optionally filtered by entity_indices

ML_100K.df2coo

Name	Type	Description
return	coo_matrix	Sparse COO matrix of shape (943, 1682) representing user-item ratings

Usage Examples

Basic Usage

from recommenders.models.geoimc.geoimc_data import ML_100K

# Initialize the ML-100K dataset handler with optional preprocessing
dataset = ML_100K(
    features_dim=50,          # Reduce features to 50 dimensions via PCA
    normalize=True,           # Normalize entity features
    target_transform="binarize"  # Binarize ratings
)

# Load data from the ML-100K directory
dataset.load_data("/path/to/ml-100k")

# Access training data
train_target = dataset.training_data.get_data()
train_user_features = dataset.training_data.get_entity(of="row")
train_item_features = dataset.training_data.get_entity(of="col")

# Access test data
test_target = dataset.test_data.get_data()
test_user_features = dataset.test_data.get_entity(of="row")
test_item_features = dataset.test_data.get_entity(of="col")