Preserve index when column_value and column_kind not provided

[jeffzi]

Hello,

This PR solves #557. The support for index based features introduced in tsfresh 0.12.0 only works when column_kind and column_value are passed to extract_features().

Here is an example of the issue:

import pandas as pd
from tsfresh.feature_extraction import extract_features
from tsfresh.feature_extraction.settings import TimeBasedFCParameters  

df = pd.DataFrame({"id": ["a", "a", "a", "a", "b", "b", "b", "b"], 
                   "value": [1, 2, 3, 1, 3, 1, 0, 8],
                   "kind": ["temperature", "temperature", "pressure", "pressure",
                            "temperature", "temperature", "pressure", "pressure"]})

# Without pd.DatetimeIndex -> empty (as expected)
extracted = extract_features(df, column_id="id", column_value='value', column_kind='kind',
                             default_fc_parameters=TimeBasedFCParameters())
extracted.info()
#> <class 'pandas.core.frame.DataFrame'>
#> Index: 0 entries
#> Empty DataFrame

# With pd.DatetimeIndex -> works
time_index = pd.DatetimeIndex(
    ['2019-03-01 10:04:00', '2019-03-01 10:50:00', '2019-03-02 00:00:00', '2019-03-02 09:04:59',
     '2019-03-02 23:54:12', '2019-03-03 08:13:04', '2019-03-04 08:00:00', '2019-03-04 08:01:00']
)
df = df.set_index(time_index).sort_index()
extracted = extract_features(df, column_id="id", column_value='value', column_kind='kind',
                             default_fc_parameters=TimeBasedFCParameters())
extracted.info()
#> <class 'pandas.core.frame.DataFrame'>
#> Index: 2 entries, a to b
#> Data columns (total 10 columns):
#> pressure__linear_trend_timewise__attr_"intercept"       2 non-null float64
#> pressure__linear_trend_timewise__attr_"pvalue"          2 non-null float64
#> pressure__linear_trend_timewise__attr_"rvalue"          2 non-null float64
#> pressure__linear_trend_timewise__attr_"slope"           2 non-null float64
#> pressure__linear_trend_timewise__attr_"stderr"          2 non-null float64
#> temperature__linear_trend_timewise__attr_"intercept"    2 non-null float64
#> temperature__linear_trend_timewise__attr_"pvalue"       2 non-null float64
#> temperature__linear_trend_timewise__attr_"rvalue"       2 non-null float64
#> temperature__linear_trend_timewise__attr_"slope"        2 non-null float64
#> temperature__linear_trend_timewise__attr_"stderr"       2 non-null float64
#> dtypes: float64(10)
#> memory usage: 176.0+ bytes

# With pd.DatetimeIndex but without specifying column_value and column_kind 
# -> empty without this PR
extracted = (df
             .groupby('kind')
             .apply(lambda g : extract_features(g, column_id="id",
                          default_fc_parameters=TimeBasedFCParameters()))
            )
extracted.info()
#> <class 'pandas.core.frame.DataFrame'>
#> MultiIndex: 0 entries
#> Empty DataFrame

^{Created on 2019-10-16 by the reprexpy package}

The bug comes from the index not being preserved when melting the dataframe in _normalize_input_to_internal_representation, specifically when column_kind and column_value are None.

[coveralls]

Coverage remained the same at ?% when pulling 23b98e0 on jeffzi:fix_index_based_features into c3d35c4 on blue-yonder:master.

[nils-braun]

Thanks for the fix.
Without further checking (sorry, not much time now, will do tomorrow): don't you end up with the additional index column?

[jeffzi]

The index doesn't appear in the results of extract_features.

The line below sets the index after the melt, therefore the index is not in the column list.

timeseries_container = timeseries_container.set_index('index')

[nils-braun]

Ah sorry sorry. I forgot that drop=True is the standard in set_index (thought it would be drop=False).
Then this is fine!

[nils-braun]

Do we also want that for the other meld in the lines above? (if you want you can also pull the two common lines with the meld an the tile out of the if branches to reduce code duplication)

[jeffzi]

Yes I think the other melt should be modified as well. I didn't test the column_sort argument before, my bad !

I've deduplicated the code as you suggested. Let me know if that's ok :)

[nils-braun]

Looks good as far as I can tell!