| title | author | institute | output | ||
|---|---|---|---|---|---|
Introduction to optimal changepoint detection algorithms |
| Toby Dylan Hocking^1^ and Rebecca Killick^2^
|
| 1. McGill University, Montreal, Canada
| 2. Lancaster University, UK
|
|
html_document |
Keywords: time series, changepoint, segmentation, optimization, machine learning.
Following the course participants will be able to:
- recognize datasets that potentially contain changepoints.
- identify appropriate changepoint methods dependent on the type of change suspected.
- create labels that indicate presence or absence of changepoints, for supervised analysis.
- perform changepoint analyses using unsupervised and supervised methods.
- check assumptions made within a changepoint analysis.
- summarize and evaluate results of a changepoint analysis.
- compare the accuracy of different labeled changepoint detection methods using cross-validation.
Changepoint detection is a part of time series data analysis which is important in fields such as finance, genomics and environment. There are many R packages that implement changepoint detection algorithms, but each package has a different interface. We will explain the differences between these packages, and show examples using several real data sets. We will emphasize coding exercises, so that participants can get familiar with using these packages. This tutorial will be ideal for useRs with time series data that do not yet know how to perform changepoint analysis in R.
Changepoint analysis is used to model time series data that has abrupt changes in the statistical distribution. The points in time when the statistical properties change are referred to as changepoints, and there are many algorithms available for computing the optimal changepoints for a given data set. This course introduces participants to changepoint analysis (also known as time series segmentation or structural change detection) and available optimal algorithms. It is highly interactive and uses packages available on CRAN and GitHub.
Techniques covered in this course include: likelihood and nonparametric methods for detecting changes in mean and variance. We will also cover supervised changepoint detection methods for labeled time series data sets. Each topic is explained theoretically and there will be break outs where participants will use the techniques on real data sets from a variety of application areas including finance, genomics and the environment.
Note that we will include interactive exercises using http://rcloud.social/, each of about 2-5 minutes. These exercises will permit students to interactively experiment and learn about changepoint detection from the R command line.
- Single and multiple changepoint detection algorithms
- Likelihood based approaches
- Exercise: Use the cpt.mean function to see if there is evidence for a change in mean in the Nile river data. If you identify a change, where is it and what are the pre and post change means?
- Exercise: Use the cpt.var function to see if there is evidence for changes in variance in the FTSE100 data from the changepoint package. If you identify changes, where are they and what are the variances in each segment?
- Exercise: Use the cpt.meanvar function to identify regions with different C+G content in the HC1 data within the changepoint package.
- Non parametric approaches
- Exercise: Look at the HeartRate data from the changepoint.np package. Use one of the non-parametric functions to see if there is evidence for changes in heart rate.
- Choosing the number of changes
- Exercise: Look at the FTSE100 data again and use the CROPS technique to determine an appropriate number of changes.
- Demonstration using changepoint and changepoint.np packages
- In changepoint detection you cannot check assumptions such as Normality prior to analysis as the changes influence any diagnostics you may perform.
- Demonstation assumption checking using changepoint package and previous class exercises.
- Exercise: Check the assumptions you have made on the simulated, Nile, FTSE100 and HeartRate data using either the segment or residual check.
- In supervised changepoint detection, there are labels which indicate presence and absence of changepoints in particular data subsets. These labels can be used for choosing the best model and parameters.
- For a given set of 6 profiles (CRAN package neuroblastoma), plot noisy data, then superimpose labels, without showing predicted changepoints.
- Exercise: plot data and labels for a different set of profiles.
- Labels can be created using prior knowledge or visual inspection.
- Exercise: create a set of labels via visual inspection for one un-labeled segmentation problem in the neuroblastoma data set. Make sure that there is at least one positive and one negative label.
- For a given labeled segmentation problem, compute optimal Gaussian changepoint models for 1 to 10 segments (CRAN package Segmentor3IsBack).
- Compute number of incorrect labels for each model.
- Choose the number of segments by minimizing the number of incorrect labels.
- Compare supervised versus unsupervised changepoint detection: many versus one data set, quantitative versus qualitative evaluation.
- Exercise: perform the same analysis on the segmentation problem that you labeled in the last section. Which models are optimal? (in terms of number of incorrect labels)
- Compute the target interval of penalty values that select changepoint models with minimal incorrect labels.
- Compute a feature vector for each segmentation problem, and a feature matrix for each labeled set of related segmentation problems.
- Learn an affine function f(feature vector)=log(penalty).
- Exercise: to learn the coefficients of the BIC penalty, what feature vector should be used?
- Un-regularized interval regression (survival package). Learns weights for a given set of features, but may overfit if non-relevant features are used.
- Elastic net regularized interval regression (anujkhare/iregnet package on GitHub). Simultaneously learns weights and performs feature selection. Avoids overfitting by setting some feature weights to zero.
- K-fold cross-validation can be used to compare prediction accuracy of supervised and unsupervised changepoint detection.
- Compute test error and ROC curves for unsupervised and supervised penalty functions: BIC, 1 feature un-regularized, multi-feature un-regularized, multi-feature regularized. Which penalty function is most accurate?
- Exercise: perform cross-validation to compare Gaussian and Logistic models for log(penalty) values. Which distribution results in more accurate penalty functions?
- Basic knowledge of R; reading in data, working with vectors and functions.
- Basic knowledge of likelihood and hypothesis testing / model choice would be useful.
We will provide a script that will automatically install all packages required for this tutorial:
- changepoint: parametric changepoint models.
- changepoint.np: non-parametric changepoint models.
- neuroblastoma: labeled data for supervised changepoint detection.
- Segmentor3IsBack: parametric changepoint models.
- survival: supervised penalty learning via un-regularized interval regression.
- iregnet: supervised penalty learning via elastic net regularized interval regression.
Time series data are rather common in many fields (finance, genomics, environment), so changepoint detection should be a rather popular topic. We expect an audience of about 50 people, so please reserve a large classroom or small lecture hall.
Toby Dylan Hocking (McGill University, Montreal, Canada; [email protected], https://github.com/tdhock) is a post-doctoral researcher, working on new machine learning models for genomic data. He has implemented several R packages for changepoint detection (neuroblastoma, bams, PeakSegDP, PeakSegJoint, coseg). He has also implemented several R graphics packages, including directlabels (won best student poster at useR2011) and animint (presented in a JSM2015 invited session and a useR2016 tutorial).
Rebecca Killick (Lancaster University, UK; [email protected], http://www.lancs.ac.uk/~killick) is a Lecturer in Statistics at the University of Lancaster. Her research is in developing methodology for the analysis of nonstationary time series to address real world problems. She has taught a range of courses over the last 10 years from first year undergraduate introductory courses to PhD level courses both theoretical and practical. She has created and contributed to several R packages on changepoint detection including changepoint, changepoint.np, EnvCpt and delivered a workshop on changepoint detection at eRum2016. Her code has been adapted for delivery in commercial software and is available as part of the NAG libraries.