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

Technion - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008
Procedure TATA-Choose-Attribute(E, A, r, ρ c )
If r = 0
Return C4.5$-Choose-Attribute(E, A, ρ c )
Foreach θ ∈ {θ ∈ A|cost(θ) < ρ c }
V ← Outcomes(θ)
Foreach vi ∈ V
Ei ← {e ∈ E | θ(e) = vi}
T ← C4.5$(Ei, A, ρ c − cost(θ))
mini ← ExpectedMC(T)
Repeat r − 1 times
T ← Stochastic-C4.5$(Ei, A, ρ c − cost(θ))
mini ← min (mini,ExpectedMC(T))
totalθ ← � |V |
i=1 mini
Return θ for which totalθ is minimal
Figure 5.4: Attribute selection in pre-contract-TATA. ExpectedMC(T) returns
the expected misclassification cost of T.
by the reconstruction process.
5.2 Contract: When Allocation is Made Right
Before Classification
The pre-contract classification scenario assumes that ρ c , the bound on testing
costs, is known to the learner. In many real-life scenarios, however, we do not
know ρ c before building the model and therefore we need classifiers that either
get ρ c as a parameter before proceeding with classification (contract classification)
or can do their best until stopped and queried for a decision (interruptible
classification). Note that TDIDT$-based algorithms cannot be used as is because
ρ c is unavailable at the time of learning. Obviously, C4.5, EG2, and ACT, can
be slightly modified, by storing default classifications at each internal node, to
produce contract and interruptible trees because they do not need the value of ρ c .
However, they are not designed to exploit a given testing budget. Therefore, we
are looking for a learner that has the advantages of pre-contract-TATA without
getting the value of ρ c as parameter.
105