Change point detection

Change point detection in time series data involves identifying moments when the underlying data-generating process shifts abruptly. This can be crucial for applications where understanding system behavior changes is essential.

1. Definition and Application: Gerrit van den Burg explains that change point detection algorithms are used to identify abrupt changes in time series data. This can help manage systems where data behavior changes sharply, such as monitoring customer numbers or in security settings. He mentions that data drift, where data distribution changes more slowly, is also related to change point detection 1 2.

2. Approaches and Algorithms: There are two main approaches to change point detection:

   • Cost-based Partitioning: Uses a cost function to minimize the number of change points detected.
   • Bayesian Methods: Learns a probability distribution over the locations of change points.

   Historical data can be used for offline detection, while continuously incoming data demands real-time or online detection algorithms 3 4.

   • 

     1

     Change Point Detection

     00:02:05
   • 

     2

     Pushing Data Limits

     00:22:08
   • 

     3

     Change Point Detection

     00:04:35
   • 

     4

     Optimizing Changepoint Detection

     00:07:05
   • 

     5

     Finding Change Points

     00:24:42
   • 

     6

     Automating Performance Testing

     00:13:05

3. Challenges: Detecting change points in real-world data can be difficult due to noise and unclear changes. In practice, defining what constitutes an abrupt change is not always straightforward, and working with noisy data adds complexity 1 3.

4. Tools and Packages: Gerrit mentions several tools for implementing change point detection, such as the SK Time package from the Alan Turing Institute for time series analysis and Ruptures for offline change point detection 2.

5. Industrial Applications: Change point detection is valuable in industrial settings for monitoring system performance and other metrics. For instance, David Daly discusses deploying these techniques in continuous integration systems to detect performance changes efficiently and reduce false positives 5 6.

For more in-depth insights, listening to the full episodes on the Data Skeptic podcast, particularly episodes on change point detection algorithms, would be beneficial.