Difference between combined_windows.json and combined_labels.json

In https://github.com/numenta/NAB/tree/master/labels, there are two types of labels: combined_windows.json and combined_labels.json. Is there a difference between the two? Thanks in advance.

1 Like

I believe that the labels file is the rawer of the two – it contains files names and a list of specific timestamps in those files where an anomaly happens.

In setting up NAB, those labeled anomalies are converted into ‘windows’ – that is, each anomaly is given a range of time surrounding it that is considered ‘in the anomaly window’. How large this window should be is configurable (and describe in the NAB paper, I believe).

For example, here is a label:

"artificialWithAnomaly/art_daily_flatmiddle.csv": [
        "2014-04-11 00:00:00"
    ],

This file has a single anomaly at "2014-04-11 00:00:00".

During setup of NAB, this one anomaly is turned into the window:

"artificialWithAnomaly/art_daily_flatmiddle.csv": [
        [
            "2014-04-10 07:15:00.000000",
            "2014-04-11 16:45:00.000000"
        ]
    ],

Any detector which detects an anomaly between the two above timestamps will have ‘detected the anomaly’ (with a higher score given for detecting the anomaly earlier in the window).