AI-Aquatica

Pipeline API

from ai_aquatica.core import WaterQualityPipeline
from ai_aquatica.datasets import load_example_dataset

data = load_example_dataset()

features = [
    "temperature",
    "pH",
    "conductivity",
    "dissolved_oxygen",
    "nitrate",
    "phosphate",
]

pipeline = (
    WaterQualityPipeline.from_dataframe(data)
    .describe()
    .ion_balance(
        cations=["Ca", "Mg", "Na", "K"],
        anions=["HCO3", "Cl", "SO4"],
        units="mg/L",
        threshold=5.0,
    )
    .select_features(features=features, target="water_quality_class")
    .impute(strategy="median")
    .scale()
    .pca(n_components=2, use_for_model=False)
    .train_random_forest(
        task="classification",
        validation="group_kfold",
        group_column="site",
        n_splits=4,
        quality_policy="warn",
        result_name="water_quality_classification",
    )
)

pipeline.export_artifacts("outputs/artifacts")
pipeline.export_html_report("outputs/report.html")

Leakage-safe preprocessing

impute(), scale(), and predictive PCA are registered but not fitted on the complete dataset. They are fitted separately inside every training partition or cross-validation fold.

PCA behaviour

• pca(..., use_for_model=False) produces exploratory scores and loadings only.
• pca(..., use_for_model=True) additionally places PCA inside the predictive scikit-learn pipeline, where it is fitted independently within every training fold.

Validation modes

• holdout
• stratified_kfold
• group_kfold
• temporal_holdout

Hydrochemical quality policies

• warn: retain non-accepted rows and emit a warning;
• filter: train only on accepted rows;
• raise: stop when non-accepted rows are present;
• ignore: do not use the status column.

API compatibility

Both .select_features(columns=[...]) and .select_features(features=[...]) are supported. The underlying processed DataFrame is available as either pipeline.data or pipeline.dataset.data.

This site is open source. Improve this page.