Machine Learning and Friends Lunch

Attribute Measurement Policies for Cost-effective Classification

Abstract

Attribute measurement is an important component of classification algorithms, which could limit their applicability in realtime settings. The time it takes to assign a value to an attribute may reduce the overall utility of the final result. I'll identify three different costs that must be considered: penalties incurred due to the misclassification of an instance, costs associated with the attribute measurement itself, and a utility function that measures the dependency of the comprehensive value on the delay in classification. I'll show how to model this problem as an MDP. I'll show how to build a policy to control this process using a set of training data and AO* heuristic search, using statistical pruning techniques and inadmissable heuristics to keep the search space manageable and avoid overfitting. I'll also discuss the problem of classifying a stream of instances, where time taken to measure attributes in the current instance can influence time-sensitive costs of waiting instances. The results offer a cost-effective approach to attribute measurement for a variety of realtime applications.

Back to ML Lunch home