A segmentation-based algorithm for large-scale partially ordered monotonic regression
2011 (English)In: Computational Statistics & Data Analysis, ISSN 0167-9473, Vol. 55, no 8, 2463-2476 p.Article in journal (Refereed) Published
Monotonic regression (MR) is an efficient tool for estimating functions that are monotonic with respect to input variables. A fast and highly accurate approximate algorithm called the GPAV was recently developed for efficient solving large-scale multivariate MR problems. When such problems are too large, the GPAV becomes too demanding in terms of computational time and memory. An approach, that extends the application area of the GPAV to encompass much larger MR problems, is presented. It is based on segmentation of a large-scale MR problem into a set of moderate-scale MR problems, each solved by the GPAV. The major contribution is the development of a computationally efficient strategy that produces a monotonic response using the local solutions. A theoretically motivated trend-following technique is introduced to ensure higher accuracy of the solution. The presented results of extensive simulations on very large data sets demonstrate the high efficiency of the new algorithm.
Place, publisher, year, edition, pages
Elsevier Science B.V., Amsterdam. , 2011. Vol. 55, no 8, 2463-2476 p.
Quadratic programming, Large-scale optimization, Least distance problem, Monotonic regression, Partially ordered data set, Pool-adjacent-violators algorithm
IdentifiersURN: urn:nbn:se:liu:diva-69182DOI: 10.1016/j.csda.2011.03.001ISI: 000291181000002OAI: oai:DiVA.org:liu-69182DiVA: diva2:424299