Installation#

pip install prefgraph

This installs the core library with NumPy, SciPy, NetworkX, Numba, and Polars. The Rust engine compiles automatically if a Rust toolchain is available. If not, a pure-Python fallback handles GARP, CCEI, MPI, and HM.

Extras#

Some workflows need additional packages. Install them with bracket syntax.

Extra	Install command	What it adds
`parquet`	`pip install "prefgraph[parquet]"`	PyArrow for reading and writing Parquet files
`datasets`	`pip install "prefgraph[datasets]"`	Pandas for loading real-world dataset loaders
`viz`	`pip install "prefgraph[viz]"`	Matplotlib for plotting
`dev`	`pip install "prefgraph[dev]"`	Pytest, Mypy, Ruff for development
`all`	`pip install "prefgraph[all]"`	Everything above plus Jupyter and Sphinx

The core pip install prefgraph is enough for load_demo, Engine, analyze_arrays, and analyze_menus. You only need extras when working with Parquet files, real-world datasets, or visualization.

Choose Your Workflow#

PrefGraph has three main entry points depending on what data you have.

I already have per-user NumPy arrays. Call Engine.analyze_arrays() directly. Each user is a tuple of (prices, quantities) arrays with shape (T, K) where T is the number of observations and K is the number of goods. See the Loading Data guide for examples.

I have a Parquet file or DataFrame. Call Engine.analyze_parquet() with column names for user ID, prices, and quantities. The engine groups by user and scores in one call. Wide format needs cost_cols and action_cols. Long format needs item_col, time_col, cost_col, and action_col. Requires pip install "prefgraph[parquet]".

I have clickstream or event logs. Build menus from your events first, then call Engine.analyze_menus(). Each user is a tuple of (menus, choices, n_items) where menus are lists of integer item indices. The Loading Data guide shows the full pipeline from raw events to scored results.