On the design of R-based scalable frameworks for data science applications

Stefan Theußl

Publikation: AbschlussarbeitDissertation

366 Downloads (Pure)


This thesis is comprised of three papers "On the design of R-based scalable frameworks for data science applications". We discuss the design of conceptual and computational frameworks for the R language for statistical computing and graphics and build software artifacts for two typical data science use cases: optimization problem solving and large scale text analysis. Each part follows a design science approach. We use a verification method for the software frameworks introduced, i.e., prototypical instantiations of the designed artifacts are evaluated on the basis of real-world applications in mixed integer optimization (consensus journal ranking) and text mining (culturomics).

The first paper introduces an extensible object oriented R Optimization Infrastructure (ROI). Methods from the field of optimization play an important role in many techniques routinely used in statistics, machine learning and data science. Often, implementations of these methods rely on highly specialized optimization algorithms, designed to be only applicable within a specific application. However, in many instances recent advances, in particular in the field of convex optimization, make it possible to conveniently and straightforwardly use modern solvers instead with the advantage of enabling broader usage scenarios and thus promoting reusability. With ROI one can formulate and solve optimization problems in a consistent way. It is capable of modeling linear, quadratic, conic, and general nonlinear optimization problems. Furthermore, the paper discusses how extension packages can add additional optimization solvers, read/write functions and additional resources such as model collections. Selected examples from the field of statistics conclude the paper.

With the second paper we aim to answer two questions. Firstly, it addresses the issue on how to construct suitable aggregates of individual journal rankings, using an optimization-based consensus ranking approach. Secondly, the presented application serves as an evaluation of the ROI prototype. Regarding the first research question we apply the proposed method to a subset of marketing-related journals from a list of collected journal rankings. Next, the paper studies the stability of the derived consensus solution, and degeneration effects that occur when excluding journals and/or rankings. Finally, we investigate the similarities/dissimilarities of the consensus with a naive meta-ranking and with individual rankings. The results show that, even though journals are not uniformly ranked, one may derive a consensus ranking with considerably high agreement with the individual rankings.

In the third paper we examine how we can extend the text mining package tm to handle large (text) corpora. This enables statisticians to answer many interesting research questions via statistical analysis or modeling of data sets that cannot be analyzed easily otherwise, e.g., due to software or hardware induced data size limitations. Adequate programming models like MapReduce facilitate parallelization of text mining tasks and allow for processing large data sets by using a distributed file system possibly spanning over several machines, e.g., in a cluster of workstations. The paper presents a plug-in package to tm called tm.plugin.dc implementing a distributed corpus class which can take advantage of the Hadoop MapReduce library for large scale text mining tasks. We evaluate the presented prototype on the basis of an application in culturomics and show that it can handle data sets of significant size efficiently.
Gradverleihende Hochschule
  • Wirtschaftsuniversität Wien
PublikationsstatusVeröffentlicht - 2020