Overview

The goal of cost-sensitive learning is to minimize data acquisition costs while maximizing the accuracy of the learner/predictor. Many fields in machine learning attempt to solve cost-sensitive learning with strong simplifying assumptions. For example, in semi-supervised learning, class-labels are assumed to be expensive and features are implicitly assumed to have zero cost. In active learning, labels are again assumed to be expensive; however the learner may ask an oracle to reveal a label for unlabeled data for selected examples. Active feature acquisition assumes that obtaining features is expensive (but typically all features are assumed to be equally expensive), and the learner identifies instances for which complete information is most informative to classify a particular test sample. Inductive transfer learning and domain adaptation methods assume that training data for a particular task is expensive or but other data from other domains may be cheaper (although relative costs are usually not explicitly modeled). Cascaded classifier architectures are primarily designed in order to reduce the cost of acquiring features to classify a sample (a sample may be classified the moment the available data is sufficient to provide sufficient classification confidence, without waiting for all features to be obtained).

There is an important but neglected common thread linking all of these different research communities. In particular, all these learning methods are motivated by the need to minimize the cost of data acquisition in many different application domains such as computer-aided medical diagnosis, computational linguistics, computational biology, and computer vision. Although all of these areas have felt the need for a principled solution to the problem, the partial solutions that have tried to solve the problem (eg semi-supervised learning, active learning, multi-task inductive transfer etc) rarely model the cost explicitly, and very little effort has been expended on modeling application specific characteristics.

Recently to some papers have started modeling the acquisition costs directly, but there is a lot of scope for theoretically rigorous work on this topic. It is also important to explicitly model the requirements from real world application communities and to bridge it with the work on theory/algorithms.

Goals

The goal of the workshop is to bring together researchers interested in the application of cost-sensitive learning (computer vision, natural language processing, computer-aided diagnostics, computational biology) with researchers interested in theory & algorithms for learning when data acquisition is costly.

The main aim is to focus attention on a practically important problem where practitioners have long sought theoretically sound algorithms but which has not been sufficiently addressed in the literature. A secondary goal is to bring together ideas from semi-supervised learning, active learning, feature acquisition, inductive transfer learning and other areas, in order that there may be more exchange of ideas across these (extremely active) communities.

Topics of Interest

The submissions on following topics are particularly encouraged (but not limited to)
Algorithms/Theory
active learning, semi-supervised learning, transfer learning, reinforcement learning, domain adaptation, cascaded classifier learning, feature selection and attribute-efficient learning, decision theory ...and related.
Applications which call for cost-sensitive learning
computer vision, computational linguistics, natural language processing, computer-aided diagnosis, differential medical diagnosis,...and others.

Important Dates

Paper submission

We call for paper contribution of up to 8 pages to the workshop using NIPS style. The accepted papers will be available for downloading from the workshop website. Accepted papers will be either presented as a talk or poster (with poster spotlight). Papers should be emailed to the organizers at cslworkshop.nips.2008@gmail.com and please indicate whether you only wish to present a poster.