Relevance based learning in Artificial Intelligence.pdf

Relevance based learning in AI

1. LEARNING USING
RELEVANT
INFORMATION

2. ALGORITHM
Hypothesis^Descriptions|= Classifications
Background^Descriptions^Classification|= Hypothesis

3. Prior knowledge: People in a country usually
speak the same language
Observation:Given nationality, language is fully
determined
Ex: Given Sima is Iranian & speaks Persian
Nat(x,n) ^Nat(y,n) ^Lang(x,l)=>Lang(y,l)
Nat(Sima, I) ^ Lang(Sima, P)

4. FUNCTIONAL DEPENDENCIES
limit the H space to
be considered
Easier to learn
target predicate
specify a sufficient basis
vocabulary to construct
hypotheses concerning the
target predicate
P⇄Q
For n Boolean features,
the determination
contains d features
DETERMINATION
(found using feature selection)
P if and only if Q

5. ADVANTAGES
Improved
learning
performance
(depends on
size of
training set)
Reduces the
required training
data
Reduce the
hypothesis
space
We combine relevance based learning with decision tree learning -> RBDTL
Time Saving
Less chance
to overfit

6. Increased accuracy
seen in RBDTL
➢A performance comparison between DECISION-TREE LEARNING and RBDTL on randomly
generated data for a target function that depends on only 5 of 16 attributes
LIMITATIONS
onoise handling
ousing other prior knowledge such as semi-supervised learning
and expert knowledge as constraints
ofrom attribute-based representation to any FOL

7. THANK
YOU!