18.11.2012 Views

anytime algorithms for learning anytime classifiers saher ... - Technion

anytime algorithms for learning anytime classifiers saher ... - Technion

anytime algorithms for learning anytime classifiers saher ... - Technion

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />

Chapter 7<br />

Conclusions<br />

The application of machine <strong>learning</strong> techniques to real world problems involves<br />

several types of costs and constraints. In this thesis we proposed a novel framework<br />

<strong>for</strong> operating under different constraints during <strong>learning</strong> and classification.<br />

Our framework allows computation speed during <strong>learning</strong> to be traded <strong>for</strong> better<br />

predictive models.<br />

In Chapter 2 we described a variety of possible scenarios of resource allocation<br />

and consumption. In contract <strong>learning</strong>, the <strong>learning</strong> resources are preallocated<br />

and provided to the learner. In interruptible <strong>learning</strong>, the learner should utilize<br />

extra time until interrupted and queried <strong>for</strong> a solution. There<strong>for</strong>e, an interruptible<br />

learner should be ready to return a valid classifier at any time.<br />

Using a classifier <strong>for</strong> predicting the labels of new cases carries two costs: the<br />

cost of the tests the model requires to administer, and the cost of the predictive<br />

errors the model makes. In cost-insensitive classification, tests do not have costs<br />

and the penalty of wrong classification is uni<strong>for</strong>m, no matter what the error type<br />

is. There<strong>for</strong>e, the objective of a <strong>learning</strong> algorithm in this case is to produce<br />

comprehensible and accurate models. In cost-sensitive classification, on the other<br />

hand, the goal is to minimize the total cost, i.e., the sum of testing costs and<br />

misclassification costs. Many real-world applications, however, limit the costs<br />

of the tests a model can require. There<strong>for</strong>e, another interesting objective we<br />

studied is to minimize misclassification costs when testing costs are bounded. The<br />

bound on testing costs may be predetermined and provided to the learner (precontract<br />

classification), known to the classifier but not to the learner (contract<br />

classification), or unknown; the classifier is then expected to exploit resources<br />

until interrupted (interruptible classification). In this thesis we handled all these<br />

scenarios.<br />

In Chapter 3 we presented a general framework <strong>for</strong> contract <strong>anytime</strong> induction<br />

of decision trees. The major limitation of top-down greedy tree learners is<br />

their inability to recover from wrong split decisions that might arise when local<br />

135

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!