Offline vs Online Change-Point Detection
What you'll learn: The fundamental difference between analyzing complete historical datasets for change-points versus detecting them in real-time as data streams in.
The Core Distinction
Once you know what change-point detection is (identifying when statistical properties shift), you need to decide when you're doing the detection.
Offline (batch) change-point detection works like a detective reviewing cold cases. You have the entire dataset already collected—perhaps months or years of measurements—and you retrospectively analyze it to find where changes occurred. You know the "ending" before you start looking for clues.
Online (real-time) change-point detection is like a security guard monitoring live camera feeds. You process data point-by-point as it arrives, deciding right now whether a change just happened, without knowing what's coming next.
Key Differences
Computational Approach
- Offline: Can use computationally intensive methods; you can look both forward and backward from any point
- Online: Must be fast enough to keep pace with incoming data; can only look backward at history
Accuracy vs Speed
- Offline: Generally more accurate because you have complete information and can refine estimates
- Online: Trade some accuracy for immediate alerts; must balance sensitivity against false alarms
Use Cases
- Offline: Quality control audits, scientific research, post-mortem failure analysis
- Online: Fraud detection systems, manufacturing monitoring, website traffic alerts
Think of it like spell-checking: offline is running a full document review after you finish writing; online is seeing red squiggles as you type.
Key Takeaway: Offline change-point detection analyzes complete datasets retrospectively with full information, while online detection must identify changes immediately as new data arrives, trading some accuracy for speed and real-time alerting capability.