Use running index as time_col about kats

[jeffhandl]

You need to specify the name of the time column if it is not time. You can do TimeSeriesData(df, time_col_name='index'). The time attribute of TimeSeriesData needs to consist of Datetime objects - how are you looking to convert the index values to Datetimes?

[orenpapers]

I am doing just that:

df :

0,0.09177
1,0.09567
2,0.08444
3,0.08971

And then I run:
TimeSeriesData(df, time_col_name='index') But then the index is automatically transformed to datetime:

                            index         0
0   1970-01-01 00:00:00.000000000  0.091773
1   1970-01-01 00:00:00.000000001  0.095667
2   1970-01-01 00:00:00.000000002  0.084443
3   1970-01-01 00:00:00.000000003  0.089711

[jeffhandl]

Yes, the times need to be Datetimes.

[orenpapers]

@jeffhandl OK , but let's say I want to treat them just as order and not specifically datetime - is it possible?

[jeffhandl]

No, you need datetime values in the time column - that is required.

[orenpapers]

@jeffhandi will be it presented in the future or there is a workaround? since ordered data is actually can be treated as a time series, even if the index is not specifically datetime

[jeffhandl]

@orko19 If you have ts_df = TimeSeriesData(df, time_col_name='index'), you can recover the indices by doing ts_df.time.index. Does that help?

[orenpapers]

@jeffhandl When I run on my dataframe:

df.apply(lambda x : CUSUMDetector(TimeSeriesData(pd.DataFrame(x['value']).
                                                                   reset_index(), time_col_name="index")).
                            detector(), axis=1)

I get:

[(TimeSeriesChangePoint(start_time: 1970-01-01 00:00:00.000000095, end_time: 1970-01-01 00:00:00.000000095, confidence: 1.0), <kats.detectors.cusum_detection.CUSUMMetadata object at 0x1db27ef40>)]
[(TimeSeriesChangePoint(start_time: 1970-01-01 00:00:00.000000057, end_time: 1970-01-01 00:00:00.000000057, confidence: 1.0), <kats.detectors.cusum_detection.CUSUMMetadata object at 0x1db27e9a0>)]
[(TimeSeriesChangePoint(start_time: 1970-01-01 00:00:00.000000229, end_time: 1970-01-01 00:00:00.000000229, confidence: 1.0), <kats.detectors.cusum_detection.CUSUMMetadata object at 0x1db27e6d0>)]
[(TimeSeriesChangePoint(start_time: 1970-01-01 00:00:00.000000095, end_time: 1970-01-01 00:00:00.000000095, confidence: 1.0), <kats.detectors.cusum_detection.CUSUMMetadata object at 0x1db27ef10>)]
[(TimeSeriesChangePoint(start_time: 1970-01-01 00:00:00.000000057, end_time: 1970-01-01 00:00:00.000000057, confidence: 1.0), <kats.detectors.cusum_detection.CUSUMMetadata object at 0x1db27e340>)]

So I meant to ask if there is a way that the output of the detector will be the actual index? Or a workaround to go back from TimeSeriesChangePoint to row/index in the original dataframe?

[jeffhandl]

Yes, let's say ts_df is the name o the time series and cp is the name of a changepoint. Then cp[0][0].start_time is the time of the changepoint, so you can do ts_df.time_to_index().get_loc(cp[0][0].start_time).