Résumé: Accurately detecting "peaks" or up changes is important in the analysis of time series and genomic data. Our ICML2015 paper describes PeakSeg, a constrained optimization model with state-of-the-art accuracy. The algorithm proposed in that paper has two drawbacks (1) it does not find the optimal solution to the PeakSeg optimization problem, and (2) its time complexity is quadratic in the number of data points. In this presentation I will show how a new functional pruning algorithm can be used to overcome both of those drawbacks. I will discuss the computational complexity (time, memory, disk space) as well as the accuracy of this new algorithm, which we have implemented in an R package.
Note: Cette présentation sera donnée en anglais. Joint work with Guillem Rigaill, Paul Fearnhead, and Guillaume Bourque
Liens: article, librairie