Transcript ICDM05bias - Columbia University

An Improved Categorization
of Classifiers’ Sensitivity on
Sample Selection Bias
Wei Fan
Ian Davidson
Bianca Zadrozny
Philip S. Yu
What is sample selection bias?


Inductive learning: training data (x,y) is sampled from the universe of
examples.
In many applications: training data (x,y) is not sampled randomly.
 Insurance and mortgage data: you only know those people you give a


policy.
School data: self-select
There are different possibilities of how (x,y) is selected (Zadrozny’04)
 S=1 denotes (x,y) is chosen.




S is independent from x and y. Total random sample.
S is dependent on y not x. Class bias
S is dependent on x not on y. Feature bias.
S is dependent on both x and y. Both class and feature.
Important Problem



It is very hard to guarantee random sample
for many real-world applications.
Heckman received Nobel Prize for his twostep approach on regression methods.
Many recent related work such as



Bianca Zadrozny’04
Andrew Smith and Charles Elkan’04.
etc
Feature Bias

P(s=1|x,y) = P(s=1|x)



Example:



Bias conditional on x
But not directly conditional on y.
Survey data
Loan approval.
Question:



Given two modeling techniques M1 and M2
Which one is more “sensitive” on feature bias?
Sensitive: constructed model and accuracy changes
significantly as a result of feature bias.
Our paper shows this



Most classifier algorithm can be sensitive or
insensitive to feature bias.
P(y|x) is the true probability distribution, which is
unknown for most problems
P(y|x,M) is the estimated probability by model M.



The dependency on M is none-trivial.
Insensitive if the model is the correct model or
asymptotically P(y|x,M) = P(y|x)
Sensitive if the model is the incorrect model or
P(y|x,M) != P(y|x)
Correct and Incorrect Model
Correct Model
Incorrect/Correct Models
Result on Decision Tree
25
20
15
Unbiased
Biased
10
5
0
1
2
3
4
5
6
Practical Implication




Given a realistic dataset, you most likely will
never know its true model either before or
after data mining.
Given a modeling technique, you will most
likely not know if it will be or will not be the
true model.
Reality is: you don’t know if it will be sensitive
or insensitive to sample selection bias.
Long paper on request.