The effectiveness of change-detection algorithms is often assessed on real-world datasets by injecting synthetically generated changes. Typically, the magnitude of the introduced changes is not controlled, and most of experimental practices lead to results that are difficult to reproduce and compare with. This problem becomes particularly relevant when the datadimension scales, as it happens with big data. To enable a fair comparison among change-detection algorithms, we have designed “Controlling Change Magnitude” (CCM), a rigorous method to introduce changes in multivariate datasets. In particular, we measure the change magnitude as the symmetric Kullback-Leibler divergence between the preand post-change distributions, and introduce changes by applying a rototranslation directly to the data. We present an algorithm to identify the parameters yielding the desired change magnitude, and analytically prove its convergence. Our experiments show the effectiveness of the proposed method and the limitations of tests run on high-dimensional datasets when changes are injected following traditional approaches. The proposed method is implemented in a MATLAB framework, which is made publicly available for download.

CCM: Controlling the Change Magnitude in High Dimensional Data

ALIPPI, CESARE;BORACCHI, GIACOMO;CARRERA, DIEGO
2016

Abstract

The effectiveness of change-detection algorithms is often assessed on real-world datasets by injecting synthetically generated changes. Typically, the magnitude of the introduced changes is not controlled, and most of experimental practices lead to results that are difficult to reproduce and compare with. This problem becomes particularly relevant when the datadimension scales, as it happens with big data. To enable a fair comparison among change-detection algorithms, we have designed “Controlling Change Magnitude” (CCM), a rigorous method to introduce changes in multivariate datasets. In particular, we measure the change magnitude as the symmetric Kullback-Leibler divergence between the preand post-change distributions, and introduce changes by applying a rototranslation directly to the data. We present an algorithm to identify the parameters yielding the desired change magnitude, and analytically prove its convergence. Our experiments show the effectiveness of the proposed method and the limitations of tests run on high-dimensional datasets when changes are injected following traditional approaches. The proposed method is implemented in a MATLAB framework, which is made publicly available for download.
Advances in Big Data
978-3-319-47897-5
978-3-319-47898-2
978-3-319-47897-5
978-3-319-47898-2
change detection
File in questo prodotto:
File Dimensione Formato  
2016_10_INNS_CCM.pdf

Accesso riservato

: Pre-Print (o Pre-Refereeing)
Dimensione 353.87 kB
Formato Adobe PDF
353.87 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11311/1001555
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 2
  • ???jsp.display-item.citation.isi??? ND
social impact