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 />
Procedure ACT-Choose-Attribute(E, A, r)<br />
If r = 0<br />
Return EG2-Choose-Attribute(E, A)<br />
Foreach a ∈ A<br />
Foreach vi ∈ domain(a)<br />
Ei ← {e ∈ E | a(e) = vi}<br />
T ← EG2(a, Ei, A − {a})<br />
mini ← Total-Cost(T, Ei)<br />
Repeat r − 1 times<br />
T ← SEG2(a, Ei, A − {a})<br />
mini ← min (mini,Total-Cost(T, Ei))<br />
totala ← Cost(a) + � |domain(a)|<br />
i=1 mini<br />
Return a <strong>for</strong> which totala is minimal<br />
Figure 4.2: Procedure <strong>for</strong> Attribute selection in ACT<br />
cost(EG2)<br />
=4.7<br />
cost(SEG2)<br />
=5.1<br />
a<br />
cost(SEG2)<br />
=4.9<br />
cost(EG2)<br />
=8.9<br />
Figure 4.3: Attribute evaluation in ACT. Assume that the cost of a in the current<br />
context is 0.4. The estimated cost of a subtree rooted at a is there<strong>for</strong>e<br />
0.4 + min(4.7, 5.1) + min(8.9, 4.9) = 10.<br />
4.3.1 Choosing a Split: Illustrative Examples<br />
ACT’s evaluation is cost-sensitive both in that it considers test and error costs<br />
simultaneously and in that it can take into account different error penalties. To<br />
illustrate this let us consider a two-class problem with mc = 100$ (uni<strong>for</strong>m) and<br />
6 attributes, a1, . . .,a6, whose costs are 10$. The training data contains 400<br />
examples, out of which 200 are positive and 200 are negative.<br />
Assume that we have to choose between a1 and a2, and that r = 1. Let<br />
the trees in Figure 4.4, denoted T1 and T2, be those sampled <strong>for</strong> a1 and a2<br />
75