Relational Sequence classification

1
Relational Sequence classification

• Part IIIin the Flf sequence trilogyby Nico
  Jacobs

2
Basic Idea

• Given a set of labeled relational strings
• Produce a readable classifier
• Underlying motivation sequence-dedicated variant
  of ILP classifier useful in e.g. user modeling
  tasks

3
The input

• List of ground atoms separated by explicit
  distance

4
The output sequences
5
When does the output cover the input?
6
Upgrade CN2

• CN2
• covering approach
• greedy beam search
• produces decision list or set of rules

7
Separate-and-conquer
8
Beam search
9
Quality criteria
10
Refinement operator
11
Unix shell data set
data(359,talk(inder), last(blob), f(blob),
f(inder), mail,sci). data(360,c, fg, fg,
man(mail), mail(camille), fg, p('kratz/proj/lates
t'), mail(kratz), f(kratz), p('kratz/proj/latest'
), head('/mbox'), head(mbox), mail, e, ls,
cd('text/651'), mail,sci). data(361,fg,talk(inde
r), f(inder), rlogin('sun-fsa'),
mail,sci). data(362,fg, mail,
rlogin('sun-fsa'), hello, ls, tidy, ll(temp), ls,
man(quota), man('8', quot), man(quot), man(du),
rm('..BAK', '..CKP'), ll, ll(temp), kill('1'),
kill('51'), ll(temp), ll(temp),ll(temp), yes(gt,
temp, ), yes, ls('.del'), quota,sci).
12
Output

• 400 sequences
• 11455 actions
• 4 groups
• majority class 25

13
Producing unordered list
14

• More rules
• More maximal distance constraints

15
Beam Size
16
Quality criteria

• No big difference between criteria
• m-estimate (with low m) best
• for unordered as well as ordered sets

17
Language

• Optimal refinement operator performs comparable
  to non-optimal operator
• add clause at front as well
• increase / decrease gaps in any order/size
• instantiate/bind vars in any order
• Should be tested on other tasks/ data sets

18
Extension Background knowledge
19
Extension Context learning
20
questions lt comments lt valuable insights