Professional Documents
Culture Documents
Distance Metric Learning
Distance Metric Learning
LEARNING
BASED ON CONVEX
FUNCTIONS PROGRAMMING
where, NM ( xi) and NH ( xi) are called the nearest miss and nearest
hit respectively.
HIT AND MISS EXAMPLES
Hit examples:
Let xi be an example in X. The hit examples of xi are the elements
of the set Hi consisting of the examples in X \ {xi} that share the
same class label with xi,
i.e. Hi = xj | j ∈ { 1 , . . . , n } , j != i, yj = yi .
Miss examples:
Let xi be an example in X . The miss examples of xi are the elements
of the set Mi consisting of the examples in X that do not share the
same class label with xi ,
i.e. Mi = xj | j ∈ { 1 , . . . , n } , yj = yi .
LARGE MARGIN DISTANCE METRIC
LEARNING APPROACH
It maximizes the margin of the nearest neighbour classifier.
D^2M(xi,xl) - d^2M(xi,xj) = [dM(xi,xl)-dM(xi,xj)] [dM(xi,xl)+dM(xi,xj)]
we can rewrite the margin based on the distance metric dM as
φM ( xi ) = dM( xi , NMm( xi )) −dM( xi , NHm( xi ) )
= gi( M ) − hi( M )
where ,gi( M ) = −min dM (xi,xj) | xj ∈ Hi
hi( M ) = −min dM (xi,xj) | xj ∈ Mi
, are convex functions of M on SD + .
IDEA FOR MISCLASSIFICATION
The idea behind the ramp loss is to truncate large losses with the constant s ,
making the classifier more robust to outliers .
CONTD..
arg min
G (M) − H(M)
𝑀≥𝑂
Essentially, the idea is to linearize the concave part and subsequently solve
the convex subproblem.