- About Anomaly Detection
- Continuous Probabilistic Methods
a. Lesson
b. Exercises
z. Miscellaneous
- Individual Project Closeout
- Overcommunicate early. Some things feel like "nothing work" so keep stakeholders informed and yourself accountable by investing in planning (e.g. kanban).
- Weekly summaries
- Daily updates
- 01_about
- The best way to detect outliers is to know your domain.
- One person's noise is another person's signal. Context is important.
- The decision rule for whether something is an inlier/outlier is usually decided by cost/benefit analysis (FP vs FN).
- Techniques for Detecting Anomalies:
- Statistical methods
- Machine learning
- Support Vector Machine
- Isolation Forest Anomal Detection
- Cluster based anomaly detection
- Density based anomlay detection
- Anomaly detection can be vulnerable to overfitting.
- 02_continuous
- Detecting anomalies in continuous data:
- Visualize
- Z score (normally distributed data)
- IQR method
- "Black Swan" events[1]
- unexpected
- severe consequences
- has a major impact
- often inappropriately rationalized after the fact with the benefit of hindsight
- 9/11
- 2008 Financial Crisis
- White Swan Event
- all of the above, except it was predicted
- COVID Pandemic
- all of the above, except it was predicted
- Grey Swan Event
- Somewhat surprising, but not completely
- Ambiguous impact (unclear what it will mean in terms of impact)
- Brexit
- Trump's election
[^1] "Black Swan Theory", Wikipedia.