anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />
a1<br />
a1<br />
a1<br />
a1<br />
a1<br />
a1<br />
a1<br />
E-MC=<br />
$110<br />
a1<br />
a1<br />
E-MC=<br />
$250<br />
a1<br />
a1<br />
E-MC=<br />
$170<br />
a1<br />
ρ c = $60<br />
Figure 5.3: Attribute evaluation in Pre-Contract-TATA. E-MC stands <strong>for</strong> the expected<br />
misclassification cost. For each candidate split, we sample the space of trees<br />
under it that fit the remaining budget ($60 in the example) and evaluate the split by<br />
the minimal expected misclassification cost in the sample ($110 in the example).<br />
linearly with r, just as it does in ACT (Esmeir & Markovitch, 2007a). When we<br />
cannot af<strong>for</strong>d sampling (r = 0), TATA builds the tree using C4.5$.<br />
5.1.4 Interruptible Learning of Pre-contract Classifiers<br />
The algorithm presented in Section 5.1.3 requires r, the sample size, as a parameter.<br />
When the <strong>learning</strong> resources are not predetermined, we would like the<br />
learner to utilize extra time until interrupted. In Chapter 3 we presented IIDT,<br />
a general framework <strong>for</strong> Interruptible Induction of Decision Trees, that need not<br />
be allocated resources ahead of time. IIDT starts with building a greedy tree.<br />
Then, it repeatedly selects a subtree whose reconstruction is expected to yield<br />
the highest marginal utility, and rebuilds the subtree with a doubled allocation<br />
of resources.<br />
The same iterative improvement approach can be applied to convert preconstract-TATA<br />
into an interruptible algorithm. The initial greedy tree would be<br />
built with C4.5$, and subtree reconstructions would be made using pre-contract-<br />
TATA. The marginal utility of constructing a tree would take into account both<br />
the expected misclassification cost of the tree and the expected resources required<br />
104<br />
a1<br />
a1<br />
a1<br />
a1<br />
The space of trees whose test cost ≤ $60