configs/ directory
src/cooprecsys/configs/
├── configuration.ini
├── fallback_config.py
├── lgbm_config.py
└── logged.pyconfiguration.ini
The shared source of defaults for progress bars, SHAP/reporting, sparse/native model defaults, DuckDB, logging, LTR-LightGBM, rating generation, paths, and the generic fallback configuration.
The most operationally important sections are:
| Section | Main use |
|---|---|
[model] | shared dtype, components, loss, schedule, epochs, native threads |
[duckdb] | DuckDB worker/thread count |
[logging] | root and tuning log levels |
[SHAP] | explanation/sample/report defaults |
[FEATURES] | LTR label/query/top-column controls |
[MODEL_LGBM] | LTR model path and base model settings |
[TRAINING] | boosting, early stopping, LambdaRank parameters |
[TUNING] | Optuna-style tuning controls |
[RATING] | pseudo-rating weights, range, method, output column |
[INFERENCE] | LTR top-k and score output settings |
[PATHS] | artifact/report/MLflow/encoder locations |
[FALLBACK] | generic fallback, ANN, cold-start, cache and execution settings |
fallback_config.py
Defines FallbackConfig, a typed dataclass that reads the [FALLBACK] section and validates options such as:
- strategy:
content,popularity,collaborative,hybrid; - score mode:
min,quantile,fixed; - ANN threshold/library/GPU flag;
- cold-start policy;
- candidate scan size and batch size;
- DuckDB and parallel execution flags.
This generic configuration class should not be confused with Ary2Tower’s dedicated TwoTowerFallBack, which is intentionally narrower and uses Bayesian-smoothed popularity with optional time decay for residual slots.
lgbm_config.py
Defines typed LightGBM configuration objects, including FeatureConfig, ModelConfig, TrainingConfig, TuningConfig, InferenceConfig, PathConfig, and the umbrella LTRConfig.
Use LTRConfig.from_ini(...) when a repeatable LTR experiment should be driven by the shared INI instead of a pile of ad-hoc command-line values.
logged.py
Provides setup_logging() so the rest of the package can use a consistent logger setup. This is especially important when debugging whether a native Cython path or a Python fallback was selected.
Configuration vs model code
Do not put large data-dependent objects into the INI file. The configuration is for defaults and policy; fitted encoders, model weights, generated reports, and large feature tables belong in artifact storage.
Example: inspect effective defaults
from pathlib import Path
from configparser import ConfigParser
cfg = ConfigParser()
cfg.read(Path('src/cooprecsys/configs/configuration.ini'))
print(cfg.get('model', 'dtype'))
print(cfg.getint('duckdb', 'threads'))
print(cfg.get('RATING', 'methodscore'))