Where academic tradition
meets the exciting future

You are here: TUCS > PUBLICATIONS > Publication Search > Machine Learning and Performan...

Machine Learning and Performance Estimation Methods for Ranking Problems

Antti Airola, Machine Learning and Performance Estimation Methods for Ranking Problems. In: Sasu Tarkoma, Joni-Kristian Kämäräinen, Tapio Pahikkala (Eds.), Proceedings of the Federated Computer Science Event 2012, 8–14, University of Helsinki, 2012.

Abstract:

The task of learning to rank refers to the machine learning problem, where the aim is to infer from past observations a ranking model that can order new objects according to how well they match some underlying criterion. Ranking problems are commonly encountered in applications such as document retrieval, game playing, information extraction and recommender systems. While learning to rank has been a topic of active research for more than a decade, developing scalable learning methods, and reliable and efficient validation methods has proven to be challenging.

The doctoral thesis of the author, summarized in this article, provides the following main contributions towards solving these issues. First, novel training algorithms based on optimizing a pairwise criterion in the
regularized risk minimization framework are derived. Previously, the most
well established method of this type is the ranking support vector machine (RankSVM). The introduced RankRLS method, as well as the proposed
improvements to RankSVM, lead to orders of magnitude gains in efficiency, without decrease in predictive performance. Second, novel cross-validation approaches are proposed in order to account for the data dependencies and multivariate performance measures characteristic of ranking tasks. Computational short-cuts allow the efficient computation of these estimates for the RankRLS method. Finally, an application study introducing a novel method for information extraction from biomedical text combines several key ideas of the thesis, resulting in a state-of-the-art solution to the problem.

Files:

Full publication in PDF-format

BibTeX entry:

@INPROCEEDINGS{inpAirola_Antti12a,
  title = {Machine Learning and Performance Estimation Methods for Ranking Problems},
  booktitle = {Proceedings of the Federated Computer Science Event 2012},
  author = {Airola, Antti},
  editor = {Tarkoma, Sasu and Kämäräinen, Joni-Kristian and Pahikkala, Tapio},
  publisher = {University of Helsinki},
  pages = {8–14},
  year = {2012},
  keywords = {cross-validation, information extraction, kernel methods, learning to rank, machine learning, regularized least-squares, regularized risk minimization, support vector machine},
}

Belongs to TUCS Research Unit(s): Turku BioNLP Group

Edit publication