Noisemaker: Splitting, Exchange & Robustness Utilities

Purpose

noisemaker/ is a data-manipulation layer. It should be used to create controlled training/evaluation conditions, not to hide business logic inside a recommender.

src/cooprecsys/noisemaker/
├── ersetz.py
├── flex.py
└── ostensible.py

flex.py: controlled train/test splitting

coo_ttsplit() handles sparse-matrix train/test splitting, while user_based_train_test_split() keeps a user-level boundary when that is required by the evaluation design.

The latter is important for recommender validation because randomly splitting individual interactions can leak a user’s preference profile into both train and validation sets.

from cooprecsys.noisemaker.flex import user_based_train_test_split

train_df, test_df = user_based_train_test_split(
    transactions,
    test_size=0.2,
    random_state=42,
)

The exact split function should match the representation used by the model and the intended production query boundary.

ersetz.py: feature/interactions exchange

This module contains utilities for transforming transaction tables into sparse interaction representations while optionally retaining user/item feature matrices and sample weights.

Its design overlaps conceptually with qrates.quasi_grade.Decomposition_Matrix_Dev(). The practical difference is intent: noisemaker is useful for controlled data manipulation/experimentation, while qrates is the scoring/rating layer.

ostensible.py: robustness transformations

extended_norm_exchange() provides another controlled transformation path around feature/interaction values. Use these helpers when testing how sensitive a model is to perturbation or alternate representations.

Production guidance

Keep the original dataset immutable and version the transformation parameters used for the experiment. A robustness experiment is only useful if another engineer can reproduce both the transformation and the baseline model result.

Avoid using noisemaker as a hidden online serving step. Serving data should pass through deterministic, explicitly versioned preprocessing instead.