Quantile regressions with lift and feed

In [1]:

!uv pip install scikit-learn lightgbm pandas pyarrow matplotlib --quiet

!uv pip install scikit-learn lightgbm pandas pyarrow matplotlib --quiet

Using Python 3.14.4 environment at: /home/niksm/code/lefts/.venv
Checked 5 packages in 4ms

In [2]:

import io
import urllib.request
import zipfile

import polars as pl
from lightgbm import LGBMRegressor

from lefts import ensemble, leaf, lift
from lefts.helpers import tabular_model
from functools import partial
import datetime as dt

import warnings

import io import urllib.request import zipfile import polars as pl from lightgbm import LGBMRegressor from lefts import ensemble, leaf, lift from lefts.helpers import tabular_model from functools import partial import datetime as dt import warnings

In [3]:

warnings.filterwarnings(
    "ignore",
    message="X does not have valid feature names, but LGBMRegressor was fitted with feature names",
)

warnings.filterwarnings( "ignore", message="X does not have valid feature names, but LGBMRegressor was fitted with feature names", )

In [4]:

url = "https://archive.ics.uci.edu/static/public/275/bike+sharing+dataset.zip"
with urllib.request.urlopen(url) as response:
    archive = zipfile.ZipFile(io.BytesIO(response.read()))
df = pl.read_csv(io.BytesIO(archive.read("hour.csv")))

df = df.with_columns(
    datetime=pl.col("dteday").str.to_datetime() + pl.duration(hours=pl.col("hr"))
).drop("dteday")

url = "https://archive.ics.uci.edu/static/public/275/bike+sharing+dataset.zip" with urllib.request.urlopen(url) as response: archive = zipfile.ZipFile(io.BytesIO(response.read())) df = pl.read_csv(io.BytesIO(archive.read("hour.csv"))) df = df.with_columns( datetime=pl.col("dteday").str.to_datetime() + pl.duration(hours=pl.col("hr")) ).drop("dteday")

In [5]:

features = ["temp", "atemp", "hum", "windspeed", "hr", "weekday", "mnth"]
target = "cnt"
quantiles = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
test_period_start_dates = pl.datetime_range(
    start=dt.datetime(2011, 3, 1),
    end=dt.datetime(2012, 12, 1),
    interval="1mo",
    eager=True,
).to_list()

quantile_models = []
for q in quantiles:
    # Convert LGBMRegressor into the format required by lefts
    base_model = leaf(
        tabular_model(
            partial(LGBMRegressor, objective="quantile", alpha=q),
            features=features,
            target=target,
        ),
        label=f"q{q}",
    )

    # 'Lift' each per-quantile model into a family of models, each with a different train and test period
    rolling_retrain = lift(
        base_model,
        name=f"q{q}_rolling_retrain",
        values=test_period_start_dates,
        # A row is in a given train period if it is
        train_filter=lambda test_period_start_date: (
            pl.col("datetime") < test_period_start_date
        ),
        test_filter=lambda test_period_start_date: (
            pl.col("datetime").dt.month() == test_period_start_date.month
        ),
        # Coalesce the outputs of the lift into a single column
        aggregate_with=pl.coalesce,
    )

    quantile_models.append(rolling_retrain)

# Stitches all the lifted models into a single ensemble that fits and predicts in parallel.
model = ensemble("quantiles", *quantile_models)

features = ["temp", "atemp", "hum", "windspeed", "hr", "weekday", "mnth"] target = "cnt" quantiles = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] test_period_start_dates = pl.datetime_range( start=dt.datetime(2011, 3, 1), end=dt.datetime(2012, 12, 1), interval="1mo", eager=True, ).to_list() quantile_models = [] for q in quantiles: # Convert LGBMRegressor into the format required by lefts base_model = leaf( tabular_model( partial(LGBMRegressor, objective="quantile", alpha=q), features=features, target=target, ), label=f"q{q}", ) # 'Lift' each per-quantile model into a family of models, each with a different train and test period rolling_retrain = lift( base_model, name=f"q{q}_rolling_retrain", values=test_period_start_dates, # A row is in a given train period if it is train_filter=lambda test_period_start_date: ( pl.col("datetime") < test_period_start_date ), test_filter=lambda test_period_start_date: ( pl.col("datetime").dt.month() == test_period_start_date.month ), # Coalesce the outputs of the lift into a single column aggregate_with=pl.coalesce, ) quantile_models.append(rolling_retrain) # Stitches all the lifted models into a single ensemble that fits and predicts in parallel. model = ensemble("quantiles", *quantile_models)

In [6]:

# You can inspect the structure of the full model before running .fit
model.print_tree()

# You can inspect the structure of the full model before running .fit model.print_tree()

Ensemble 'quantiles' (198 models)  → outputs: [q0.1_rolling_retrain, ..., q0.9_rolling_retrain]
    ├── Lift 'q0.1_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.1_rolling_retrain"
    │   └── Leaf 'q0.1'
    ├── Lift 'q0.2_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.2_rolling_retrain"
    │   └── Leaf 'q0.2'
    ├── Lift 'q0.3_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.3_rolling_retrain"
    │   └── Leaf 'q0.3'
    ├── Lift 'q0.4_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.4_rolling_retrain"
    │   └── Leaf 'q0.4'
    ├── Lift 'q0.5_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.5_rolling_retrain"
    │   └── Leaf 'q0.5'
    ├── Lift 'q0.6_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.6_rolling_retrain"
    │   └── Leaf 'q0.6'
    ├── Lift 'q0.7_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.7_rolling_retrain"
    │   └── Leaf 'q0.7'
    ├── Lift 'q0.8_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.8_rolling_retrain"
    │   └── Leaf 'q0.8'
    └── Lift 'q0.9_rolling_retrain' (22 models): [2011-03-01 00:00:00, ..., 2012-12-01 00:00:00]  ⇒ coalesce → "q0.9_rolling_retrain"
        └── Leaf 'q0.9'

In [7]:

# Fits |quantiles| x |test_period_start_dates| = 9 * 22 = 198 models
model.fit(df, logging="drop", errors="raise")

# Fits |quantiles| x |test_period_start_dates| = 9 * 22 = 198 models model.fit(df, logging="drop", errors="raise")

In [8]:

# Adds |quantiles| columns, each with the unique prediction associated with that test row.
predictions = model.predict(df, errors="raise")

# Adds |quantiles| columns, each with the unique prediction associated with that test row. predictions = model.predict(df, errors="raise")

In [10]:

predictions.filter(pl.col("datetime").dt.date() == dt.date(2012, 12, 1)).select(
    "datetime", *model.collect_labels()
).to_pandas().set_index("datetime").plot(figsize=[15, 4], alpha=0.3)

predictions.filter(pl.col("datetime").dt.date() == dt.date(2012, 12, 1)).select( "datetime", *model.collect_labels() ).to_pandas().set_index("datetime").plot(figsize=[15, 4], alpha=0.3)

Out[10]:

<Axes: xlabel='datetime'>

In [9]: