Professional Documents
Culture Documents
2 iI(j) w2
P
+
2( )2
j Pe (j) = P(qj 6= qj ) e ij . (5)
P 2
1e iI(j) (bi ai )
Replacing the weight wij by 2pij 1, we have
( iI(j) (2pij 1))2
P
where wij distributes in the range [ai , bi ] = [wij , wij ] uni-
Pe (j) = P(qj 6= qj ) e 2 . (6)
formly [Hoeffding, 1963]1 . Similarly, we can derive that the
expected accuracy rate of WMV if the true value of j is -1 And the expected accuracy rate P(qj = qj ) of each assign-
ment j by the WMV aggregation method satisfies
We will buy additional pages to include the appendix in the final Pa (j) = P(qj = 1 Pe (j) (7)
version if the paper is accepted. P 2
( iI(j) (2pij 1))
1
To make the Hoeffding inequality holds, the value of j+ 0
and in the latter weight setting, we will ensure this non-negative
= qj ) 1 e 2 . (8)
condition holds.
2 Proof of Theorem 3 j H, Pue (j, y)=Pue (j, x)), where each assignment in H
Submodular Property: Let S1 and S2 denote two diligent is graded diligently by the peers under both of x and y.
peer-assignment pair subsets, where S1S2 . For any diligen- Ineq.(10) holds because 1) j H, Pue (jH , x) Pue (j, x)
u
t peer-assignment grading (i , j ) S / 2 , S1 =S1 (i , j ) and and xj yj and 2) and j L, Pe (jL , x) Pue (j, x) and
S2 =S2 (i , j ), then, we have we proceed to compute xj yj . Eq.(11) holds because Pue (jL , y)Pue (jH
, y), where
the assignment jH that is fully diligently graded by the peers
n (S1 ) n (S1 ) (n (S2 ) n (S2 )) has a smaller expected error rate than expected error rate of
S
= Pue (j , I(j , S1 )\{i }) (1xSj 1 )(1xj 1 )Pue (j , {i })
the assignment jL that might be partially diligently graded
by the peers. Formally, denote jH s expected error rate of
S
Pue (j , I(j , S2 )\{i }) (1xSj 2 )(1xj 2 )Pue (j , {i })
(2p1 1)2
P
0.5
fully diligent grading as Pue (jH , x)=e iI(j )
H
i
,
S
= Pue (j , I(j , S1 )\{i }) (1xSj 1 )(1xj 1 )Pue (j , {i })
where |I(jH )|=l, and jL s expected error rate of partially
(2p1i 1)2
P
0.5 0
Pue (j , I(j , S2 \S1 )\{i }) (1xSj 2 ) diligent grading as Pue (jL , x)=e iI (j )
L , where
0 0 0
S
I (jL ) I(jL ) and |I (jL )|=l l. Because each assign-
(1xj 2 )Pue (j , {i })
ment is randomly allocated to the peers with probability n1 ,
S then we have jH s expected error ratePof fully diligent grad-
(1 xSj 1 ) (1 xj 1 )Pue (j , {i }) l 1 2
ed by l peers is Pue (jH
, x)=e0.5 n iI (2pij 1) and jL s
Pue (j , I(j , S2 )\I(j , S1 ) \ {i }) 0
expected error rate of fully diligent graded by l peers is
S 0
(1xSj 2 )(1xj 2 )Pue (j , {i })
l
P 1 2 0
Pue (jL , x)=e0.5 n iI (2pi 1) . Because l l, we have
(1xSj 1 )(1x
S1
)Pue (j , {i })
Pue (jL , x)Pue (jH , x).
j P
Case 2), If K > j j . Formally, resort the assignments
S
(1xSj 2 )(1xj 2 )Pue (j , {i })
in decreasing order of their error rate under fully diligen-
S S t grading, i.e., Pue (1) Pue (2) . . . Pue (n), let x={x1 =
= (xSj 2 xSj 1 )Pue (j , {i })(xj 2 xj 1 ) 1 + xr1 , x2 = 2 + xr1 , . . . , xn = nP + xrn } denote Algorith-
S S S S
(xSj 2 xSj 1 )(xj 2 xj 1 ) = (xj 1 xSj 1 )(xj 2 xSj 2 )0.
r r
m 1s policy, where xj =max{K 1q<j xrq , 1 j } and
K r =K j j is the remaining budget after diligent elic-
P