Loading...

Proactive Learning for Solving Classification Problems

Hafez Kolahi, Hassan | 2013

1152 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 45146 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Rabiee, Hamid Reza
  7. Abstract:
  8. The goal of active learning is to choose the training instances intelligently, in order to make the cost of the learning process as low as possible. In proactive learning, which is an extension of active learning, several noisy labelers provide labels for each chosen sample. The proactive problem consists of three different subproblems: estimating true labels and finding their correctness probability (crowdsourcing), learning a model for finding labels of new instances (classification), and selecting samples for empowering the classifier (active learning). Since estimated labels through crowdsourcing are not certain, regular classification and active learning methods should be extended to be able to use labels’ uncertainty. Considering good performance of Support Vector Machines (SVM), we introduce Uncertain Input SVM (UISVM) to solve the target problem. To the best of our knowledge, our work is the first to propose such an extension for SVMs. In this thesis, we discuss challenges of using uncertainty in SVM and provide solutions to overcome them. We also show that modeling the UISVM as a parametric quadratic programming problem allows us to use some results from parametric optimization field.Using some of these theories, we provide an efficient algorithm for query selection, based on maximum expected model change. Finally, we provide empirical results showing the superiority of the proposed algorithm over methods proposed in the literature
  9. Keywords:
  10. Active Learning ; Proactive Learning ; Crowdsourcing ; Crowd Computing

 Digital Object List

 Bookmark

...see more