Neighbors

A Python package for collaborative filtering on social datasets

Installation

1. Pip (official releases): pip install neighbors
2. Github (bleeding edge): pip install git+https://github.com/cosanlab/neighbors.git

Getting started

The best way to learn how to use the package is by checking out the documentation site which contains usage tutorials as well as API documentation for all package functionality.

Quick Demo Usage

from neighbors.models import NNMF_sgd
from neighbors.utils import create_user_item_matrix, estimate_performance

# Assuming data is 3 column pandas df with 'User', 'Item', 'Rating'
# convert it to a (possibly sparse) user x item matrix
mat = create_user_item_matrix(df)

# Initialize a model
model = NNMF_sgd(mat)

# Fit
model.fit()

# If data are time-series optionally fit model using dilation
# to leverage auto-correlation and improve performance
model.fit(dilate_by_nsamples=60)

# Visualize results
model.plot_predictions()

# Estimate algorithm performance using
# Repeated refitting with random masking (dense data)
# Or cross-validation (sparse data)
group_results, user_results = estimate_performance(NNMF_sgd, mat)

Algorithms

Currently supported algorithms include:

• Mean - a baseline model
• KNN - k-nearest neighbors
• NNMF_mult - non-negative matrix factorization trained via multiplicative updating
• NNMF_sgd - non-negative matrix factorization trained via stochastic gradient descent

By default both NNMF models clip their predictions to the range of the observed ratings, since matrix factorization (and especially the unconstrained bias terms in NNMF_sgd) can otherwise produce predictions outside that range (e.g. negative values despite all-positive ratings). This is the same approach the Surprise package takes when making predictions, and can be disabled with .fit(clip_predictions=False). If the observed ratings don't span the full rating scale (e.g. no one used the top of a 1-5 scale), pass the scale explicitly with .fit(clip_range=(1, 5)).