irspack - Implicit recommenders for practitioners#

Build, compare, and tune recommenders from clicks, views, saves, and purchases. irspack provides a consistent workflow for taking implicit-feedback data from a practical baseline to an evaluated model candidate.

Prepare your data

Convert business IDs and event tables into sparse interaction matrices.

Using your own interaction data
Choose a model

Start with a useful baseline and compare models with different biases.

Choosing a recommender
Evaluate consistently

Compare candidates with the same split, candidate set, and ranking metrics.

Evaluation of recommender systems
Tune promising models

Use Optuna-backed tuning after a model earns further experimentation.

Hyperparameter Optimization

Why irspack?#

There is no all-purpose algorithm for implicit-feedback recommendation. The practical approach is to establish a baseline, compare candidates under one evaluation protocol, and tune the models that work for your data. irspack keeps that workflow consistent across algorithms.

Key capabilities include:

  • optuna-backed, efficient hyperparameter optimization. In particular, pruning is used to speed-up the parameter search for several algorithms.

  • Implementation of several parallelizable algorithms with nanobind and Eigen. Evaluation of recommenders’ performance (which involves score-sorting and ranking metric computation) can be also done efficiently using these technologies.

Install and start#

Install the published package from PyPI:

pip install irspack

Then begin with Using your own interaction data, or work through the first Train our first movie recommender notebook.

Pre-built wheels are published for supported Linux, macOS, and Windows platforms. To build with CPU-specific compiler optimizations, install from source with:

CFLAGS="-march=native" pip install git+https://github.com/tohtsky/irspack.git

Reference indices#