Mangaki's recommendation algorithms.
They are tested on Python 3.8, 3.9, 3.10 over OpenBLAS LP64. Intel MKL is broken for now, see #25.
poetry install
To run cross-validation:
- Download a dataset like Movielens 100k.
- Ensure the columns are named
user,item,rating:
| user | item | rating |
|---|---|---|
| 3 | 5 | 4.5 |
For example, here, user 3 gave 4.5 stars to item 5.
-
Then run:
python compare.py <path/to/dataset>
You can tweak the experiments/default.json file to compare other models.
This assumes you installed the dependencies of the project.
- To run tests for the Python Zero module:
py.test zero - To run tests for the Aggregation module:
py.test aggregation
Most models have the following routines:
from zero.als import MangakiALS
model = MangakiALS(nb_components=10)
model.fit(X, y)
model.predict(X)
where:
- X is a numpy array of size
nb_samplesx 2 (first column: user ID, second column: item ID) - and y is the column of ratings.
There are a couple of other methods that can be used for online fit, say model.predict_single_user(work_ids, user_parameters).
See zero.py as an example of dumb baseline (only predicts zeroes) to start from.
Install this module with the extra secure-aggregation, i.e. pip install mangaki-zero[secure-aggregation] or compile the module in aggregation/, this only requires a stable Rust compiler (CI tests are performed against Rust stable, beta and nightlies.) and maturin.
Then, you can follow the docs there: https://mangaki.github.io/zero/
Feel free to use. Under MIT license.