Investigate unusual readings in Orange

A shop closes for the day, yet its meter records more electricity than on any open day. In our invented example, that reading needs a question beside it: was equipment running, was the meter period different, or was a number copied incorrectly? The first useful result is a shortlist to investigate.

Isolation Forest repeatedly divides data using a randomly chosen feature and split value. Observations that separate after fewer divisions, averaged across many such trees, receive more anomalous scores. It learns the shape of the supplied numbers without needing examples labelled “wrong”.[5]

This documentation-based exercise was checked on 29 October 2026; we did not run the Orange interface or observe its selections.

Use Orange 3.40.0 desktop for Windows or macOS.[1]

Practice log

Original practice data: each row is one day. The columns are ID, hours open, and daily electricity consumption in kWh.

• A: 8 hours, 24 kWh
• B: 8 hours, 25 kWh
• C: 7 hours, 21 kWh
• D: 9 hours, 27 kWh
• E: 8 hours, 23 kWh
• F: 7 hours, 22 kWh
• G: 9 hours, 28 kWh
• H: 0 hours, 5 kWh
• I: 0 hours, 6 kWh
• J: 0 hours, 35 kWh

Enter these ten rows in a spreadsheet using column names ID, hours, kWh, then save a CSV. Keep the original sheet. The exercise uses only opening hours and electricity consumption; the letter identifies the day so you can find its original record.

Make the shortlist

On Orange’s canvas, add and connect File to Outliers. Load the CSV. In File, double-click columns: hours/kWh → Continuous, Features; ID → String, Meta. Keep dates/categories in Meta.[2]

In Outliers choose Isolation Forest, Contamination 10%, Replicable training; Apply. Connect Outliers’ Outliers output to Save Data. Save as CSV, unticking Add type annotations to header. Repeat at 20%, saving separately.[3][4]

Read both exports

Contamination sets the threshold used to turn scores into an outlier selection. Raising it widens the requested share of unusual observations; it does not discover the percentage of erroneous readings. The cutoff is a choice made before reviewing the results.[5]

Open both CSVs and compare their IDs with the original sheet. What appeared only with the wider cutoff? For our invented log, J deserves inspection because its 35 kWh accompanies zero opening hours; that is our reading of the log, not a reported software result. A flagged closed day could also be perfectly legitimate.

If every row appears, check the connection uses Outliers rather than Data.[3]

Write a review note beside each selected ID while preserving the measured values. The useful finish is a set of traceable questions, with the cutoff recorded. This tiny log rehearses the workflow; a real shop would need a representative history of its ordinary open and closed days before judging how useful the shortlist is.

Report an error or suggest a correction