Information Technology Reference
In-Depth Information
most similar cluster
SC
will define the expertise based weighting function's pa-
rameters. Given these notations, the fused score for a certain instance
E
will be
calculated as follows:
FusedScore(E)=s*[average score of expertsADAs(SC)]+(1-s)*[average score
of unexpertsADAs(SC)]
5 Preliminary Experiment
In this section we limit ourselves to the case of a single outlier type. For this case
we describe the Union Voting Fusion method that basically aims to overcome
the normalization issue.
5.1 The Union Voting Fusion Algorithm
The scores from each ADA are converted to rankings. Then these rankings are
combined into a (inverse) suspicious score by taking the k-th highest rank from
the group of the ADAs. This is interpreted as at least k ADAs having a consensus
that the final rankings of the data are not too suspicious. For example, if an
instance's final suspicious score is 10, then k of the classifiers agree that this
instance deserves to be on a top-10 outlier list.
It is preferable to combine rankings rather than the raw scores due to the fact
that the scores cannot be interpreted in the same way. The rankings obtained in
this manner are more robust than the individual rankings of the outlier detectors
because they are smoothed out.
Tabl e 1.
Scores table
Instance Score 1
Score 2
Score 3
0.5 (1
st
) 0.75 (3
rd
) 0.9 (1
st
)
A
0.4 (2
nd
) 0.9 (2
nd
) 0.8 (3
rd
)
B
0.3 (3
rd
) 0.95 (1
st
) 0.85 (2
nd
)
C
0.2 (4
th
) 0.3 (4
th
) 0.2 (5
th
)
D
0.1 (5
th
) 0.1 (5
th
) 0.25 (4
th
)
E
Example.
If we do a “one-vote union”, we take the smallest ranking as the
ranking of the 3-detector ensemble. i.e.
A : min(1,3,1) = 1
B : min(2,2,3) = 2
C : min(3,1,2) = 1
...
and so on
Instances with the score
n
appeared in the top-
n
outlier list of at least 1 detec-
tor. In other words the “one-union” means that it is sucient to be considered
by at least one ADA classifier.
Search WWH ::
Custom Search