[BUG] Feature selection transforms fail to make inverse_transform after make_future

[Mr-Geekman]

🐛 Bug Report

Feature selection transforms with option return_features=True work incorrectly during inverse_transform.

During transform it stores all the removed features inside transform and returns it back during inverse_transform. The problem occurs when we try to make inverse_transform after make_future because after this operation our dataset become smaller and concatenation with data saved during transform corrupts the result.

I think that we shouldn't store any state during fit. As I understand, this problem can be solved if we only hide this columns in TSDataset instead of removing and unhide them during inverse_transform. This makes the task quite difficult.

Expected behavior

Feature selection transforms works fine during inverse_transform after make_future.

How To Reproduce

Script with bug:

import pandas as pd
import numpy as np

from etna.transforms import FilterFeaturesTransform
from etna.datasets import TSDataset


def get_regular_ts() -> TSDataset:
    periods = 100
    df_1 = pd.DataFrame({"timestamp": pd.date_range("2020-01-01", periods=periods)})
    df_1["segment"] = "segment_1"
    df_1["target"] = np.random.uniform(10, 20, size=periods)

    df_2 = pd.DataFrame({"timestamp": pd.date_range("2020-01-01", periods=periods)})
    df_2["segment"] = "segment_2"
    df_2["target"] = np.random.uniform(-15, 5, size=periods)

    df_3 = pd.DataFrame({"timestamp": pd.date_range("2020-01-01", periods=periods)})
    df_3["segment"] = "segment_3"
    df_3["target"] = np.random.uniform(-5, 5, size=periods)

    df = pd.concat([df_1, df_2, df_3]).reset_index(drop=True)
    df = TSDataset.to_dataset(df)
    tsds = TSDataset(df, freq="D")

    return tsds


def get_ts_with_exog() -> TSDataset:
    regular_ts = get_regular_ts()
    df = regular_ts.to_pandas(flatten=True)
    df_exog = df.copy().drop(columns=["target"])
    df_exog["weekday"] = df_exog["timestamp"].dt.weekday
    df_exog["monthday"] = df_exog["timestamp"].dt.day
    df_exog["month"] = df_exog["timestamp"].dt.month
    df_exog["year"] = df_exog["timestamp"].dt.year
    ts = TSDataset(df=TSDataset.to_dataset(df).iloc[5:-20], df_exog=TSDataset.to_dataset(df_exog), freq="D")
    return ts


def main():
    ts = get_ts_with_exog()
    transform = FilterFeaturesTransform(exclude=["year"], return_features=True)

    ts.fit_transform([transform])
    future_ts = ts.make_future(future_steps=10, tail_steps=0)

    future_ts.inverse_transform()
    future_df = future_ts.to_pandas()

    assert len(future_df) == 10


if __name__ == "__main__":
    main()

This script fails because in reality we have len(future_df) == 85.

Environment

No response

Additional context

No response

Checklist

• Bug appears at the latest library version