Explaining a Result to the End-User: A Geometric Approach for

It gives global indicators about the easiness to maintain viability. It gives information about the resilience r(u) of trajectories. Green trajectories are more resilient.
486KB taille 1 téléchargements 133 vues
Explaining a Result to the End-User: A Geometric Approach for Classification Problems I.Alvarez & S.Martin Geometric information identifies in classification trees the most relevant tests to describe the situation (Alvarez, 04)

Geometric approach gives information about the robustness of the decision considering change in the input case

Decision tree for next year course applicants H1: Mark in mathematics is less than 50/100 yes

no

robustness r of case x r = d(x, Γ)

H2: Mean is less than 50/100

Repeat the year Andrew Math: 40/100 Mean: 30/100

yes

Repeat the year

Pass

Andrew’s trace of classification: mark in math < 50

Pass

B

50

p(Andrew) Andrew

H2

Fig. 1: The trace of Andrew’s classification doesn’t describe the situation (case B is still in the same class area), contrary to the tests defining case p(Andrew)

50

oligotrophic lake

eutrophic lake

The state of the lake is defined by the amount of phosphorus and by phosphorus inputs from agriculture (Carpenter et al, 99)

mean

Geometric information presents a complementary viewpoint to probabilistic information. Distant points from the decision

B(x,r)

Application to a deterministic classifier for the eutrophication lake problem

loss P q (t ) recycling from dP (t ) = − b.P(t ) + L(t ) + r. q m + P q (t ) sediments dt

Regulation law is a constraint on dL/dt. The viability model gives the set of states where it is possible to maintain the oligotrophic state and agriculture (Martin, 05) 100

boundary with low probability

1.4

Class membership probability 1.0

4

B

σ1

3 σ2

Γ

1.0

0.8

A

2

V2

1.2

0.8

p(A)

0.6

p(B)

1

P

math

Math: 51/100 Mean: Andrew’s mean

decision boundary Γ

x

no

Green trajectories are more resilient

0

u1

0.6

u2

Fig. 3:The distance to the viability decision boundary Γ gives infirmation about the robustness of viable states. It gives global indicators about the easiness to maintain viability. It gives information about the resilience r(u) of trajectories.

p(A )

0

0.4

A

rh (u ) = min t ≤ h {d (u (t ), Γ )}

0.4

-1

B

0.2

p(B)

M’

0.2

-2 M

0.0 -2

-1

0

1

2

3

4

V1

Fig. 2: Points with the same class membership probability can have a very different geometric situation. A is close to the decision boundary, a small change in its attribute values can change the decision.

0.0

-100 0.0

0.2

0.4

0.6 L

0.8

Algebraic Distance (%max)

Large areas inside the viability set are good indicator for viability control

Limits: Geometric approach applies to metric space x − min i only. Basic metrics: min − max y = i i

Computation costs can be exponential (Meijster et al, 00). standard

yi =

max i − min i xi − xi

σi