A library for preparing and evaluating scalable deep learning hybrid recommender systems

collie

A library for preparing, training, and evaluating scalable deep learning hybrid recommender systems using PyTorch.

Collie is a library for preparing, training, and evaluating implicit deep learning hybrid recommender systems, named after the Border Collie dog breed.

Collie offers a collection of simple APIs for preparing and splitting datasets, incorporating item metadata directly into a model architecture or loss, efficiently evaluating a model's performance on the GPU, and so much more. Above all else though, Collie is built with flexibility and customization in mind, allowing for faster prototyping and experimentation.

Installation

pip install collie

Through July 2021, this library used to be under the name collie_recs. While this version is still available on PyPI, it is no longer supported or maintained. All users of the library should use collie for the latest and greatest version of the code!

Quick Start

Implicit Data

Creating and evaluating a matrix factorization model with implicit MovieLens 100K data is simple with Collie:

from collie.cross_validation import stratified_split
from collie.interactions import Interactions
from collie.metrics import auc, evaluate_in_batches, mapk, mrr
from collie.model import MatrixFactorizationModel, CollieTrainer
from collie.movielens import read_movielens_df
from collie.utils import convert_to_implicit


# read in explicit MovieLens 100K data
df = read_movielens_df()

# convert the data to implicit
df_imp = convert_to_implicit(df)

# store data as ``Interactions``
interactions = Interactions(users=df_imp['user_id'],
                            items=df_imp['item_id'],
                            allow_missing_ids=True)

# perform a data split
train, val = stratified_split(interactions)

# train an implicit ``MatrixFactorization`` model
model = MatrixFactorizationModel(train=train,
                                 val=val,
                                 embedding_dim=10,
                                 lr=1e-1,
                                 loss='adaptive',
                                 optimizer='adam')
trainer = CollieTrainer(model, max_epochs=10)
trainer.fit(model)
model.eval()

# evaluate the model
auc_score, mrr_score, mapk_score = evaluate_in_batches(metric_list=[auc, mrr, mapk],
                                                       test_interactions=val,
                                                       model=model)

print(f'AUC:          {auc_score}')
print(f'MRR:          {mrr_score}')
print(f'MAP@10:       {mapk_score}')

More complicated examples of implicit pipelines can be viewed for MovieLens 100K data here, in notebooks here, and documentation here.

Explicit Data

Collie also handles the situation when you instead have explicit data, such as star ratings. Note how similar the pipeline and APIs are compared to the implicit example above:

from collie.cross_validation import stratified_split
from collie.interactions import ExplicitInteractions
from collie.metrics import explicit_evaluate_in_batches
from collie.model import MatrixFactorizationModel, CollieTrainer
from collie.movielens import read_movielens_df

from torchmetrics import MeanAbsoluteError, MeanSquaredError


# read in explicit MovieLens 100K data
df = read_movielens_df()

# store data as ``Interactions``
interactions = ExplicitInteractions(users=df['user_id'],
                                    items=df['item_id'],
                                    ratings=df['rating'])

# perform a data split
train, val = stratified_split(interactions)

# train an implicit ``MatrixFactorization`` model
model = MatrixFactorizationModel(train=train,
                                 val=val,
                                 embedding_dim=10,
                                 lr=1e-2,
                                 loss='mse',
                                 optimizer='adam')
trainer = CollieTrainer(model, max_epochs=10)
trainer.fit(model)
model.eval()

# evaluate the model
mae_score, mse_score = explicit_evaluate_in_batches(metric_list=[MeanAbsoluteError(),
                                                                 MeanSquaredError()],
                                                    test_interactions=val,
                                                    model=model)

print(f'MAE: {mae_score}')
print(f'MSE: {mse_score}')

Comparison With Other Open-Source Recommendation Libraries

On some smaller screens, you might have to scroll right to see the full table. ➡️

The following table notes shows the results of an experiment training and evaluating recommendation models in some popular implicit recommendation model frameworks on a common MovieLens 10M dataset. The data was split via a 90/5/5 stratified data split. Each model was trained for a maximum of 40 epochs using an embedding dimension of 32. For each model, we used default hyperparameters (unless otherwise noted below).

At ShopRunner, we have found Collie models outperform comparable LightFM models with up to 64% improved MAP@10 scores.

Development

To run locally, begin by creating a data path environment variable:

# Define where on your local hard drive you want to store data. It is best if this
# location is not inside the repo itself. An example is below
export DATA_PATH=$HOME/data/collie

Run development from within the Docker container:

docker build -t collie .

# run the container in interactive mode, leaving port ``8888`` open for Jupyter
docker run \
    -it \
    --rm \
    -v "${DATA_PATH}:/collie/data/" \
    -v "${PWD}:/collie" \
    -p 8888:8888 \
    collie /bin/bash

Run on a GPU:

docker build -t collie .

# run the container in interactive mode, leaving port ``8888`` open for Jupyter
docker run \
    -it \
    --rm \
    --gpus all \
    -v "${DATA_PATH}:/collie/data/" \
    -v "${PWD}:/collie" \
    -p 8888:8888 \
    collie /bin/bash

Start JupyterLab

To run JupyterLab, start the container and execute the following:

jupyter lab --ip 0.0.0.0 --no-browser --allow-root

Connect to JupyterLab here: http://localhost:8888/lab

Unit Tests

Library unit tests in this repo are to be run in the Docker container:

# execute unit tests
pytest --cov-report term --cov=collie

Note that a handful of tests require the MovieLens 100K dataset to be downloaded (~5MB in size), meaning that either before or during test time, there will need to be an internet connection. This dataset only needs to be downloaded a single time for use in both unit tests and tutorials.

Docs

The Collie library supports Read the Docs documentation. To compile locally,

cd docs
make html

# open local docs
open build/html/index.html