How to read these slides?
Overview
Click on the menu bar items to navigate to chapters
Click here for PDF version (Note: No videos in the PDF version)
AnoScout
Visual Exploration of Anomalies and Anomaly Detection Algorithm Ensembles in Time Series Data
Julian Rakuschek, Michael Leitner, Jürgen Bernard, Selina C. Wriessnegger, Tobias Schreck
VINCI 2025 - 01.12.2025
Anomalies in time series (Selection)
Anomaly = Unexpected Pattern
Easy to see for a human!
Check every time series by hand?
Let algorithms do the work!
We would like to find these anomalies:
Apply unsupervised algorithms = no training required
Unsupervised algorithms produce a scoring
Input Time Series
K-Means
Better for first anomaly
Local Outlier Factor
Better for second anomaly
Ensemble to combine strengths
Input Time Series
Ensemble
Average of LOF and K-Means
Threshold
Everything above threshold = anomaly
The Result are Interval Annotations for Anomalies
Not perfect, but close enough
Coming back to this:
We now ask:
- Which anomalies arise?
- Anomaly categorization?
- Strengths and weaknesses of algorithms?
Introducing AnoScout
AnoScout = Playground for anomaly detection algorithms
Our contribution: A workflow to explore anomalies
Scenario: EEG Artifact Detection
Extract Anomalies
Each anomaly is represented through a card
Explore Anomalies
Projection Based
Step 3: Rank Anomalies
This was all unsupervised.
What if I already know the expected behavior?
Scenario: Manufacturing
How can we configure a classifier for that?
AnoScout supports configuring a classifier
Explore Anomalies
Cluster Based
Summary
Future Work
- Algorithm Parameter Guidance
- Multivariate Time Series
- Progressive Exploring for Large Datasets
Thank you!
Julian Rakuschek
Michael Leitner
Jürgen Bernard
Selina Wriessnegger
Tobias Schreck
Open Source
Slides