Anomaly Detection Task output should include PredictedLabel

[colbylwilliams]

The Inputs and Output Columns section should have PredictedLabel in addition to Score. Note: There is an issue with the current release of ML.NET where PredictedLabel always evaluates to true.

Additionally, the description for Score reads:

The non-negative, unbounded score that was calculated by the anomaly detection model.

This does not indicate how the value returned by the model for Score should be interpreted. Should higher scores be interpreted as anomalies, or scores closer to zero?

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: 8b318bcc-48ff-eb63-9952-e811a283eb90
• Version Independent ID: 1810a47c-da4c-8a3a-f5a0-61b31069f083
• Content: RandomizedPcaTrainer Class (Microsoft.ML.Trainers)
• Content Source: dotnet/xml/Microsoft.ML.Trainers/RandomizedPcaTrainer.xml
• Product: dotnet-ml-api
• GitHub Login: @sfilipi
• Microsoft Alias: johalex

[colbylwilliams]

Confirmed here, Predictions that are anomalies will have higher scores than those that aren't anomalies.

What it is essentially computing is 1 minus the ratio of the length of the projected vector and the input vector. We expect the higher the scores the higher the chance of an anomaly because in case of no anomaly the data point should be able to be projected without changing the length much.