Download Pattern Directed Mining Of Sequence Data

Pattern Directed Mining Of
Sequence Data
By Valery Guralnik,
Duminda Wijesekera,
Jaideep Srivastava
Presenter : Jyothsna R
Nayak
contents
Introduction
Sequential Patterns
Data Structure and Algorithm
Experimental Evaluation
SP Tree Optimization
Conclusions
References
Introduction
Sequence data
event has an associated time of
occurrence
Episode is a collection of events
Frequent Episodes : Episodes occurring
with a frequency above a certain
threshold
Steps involved in mining of
frequent episodes
Present a language for specifying episodes of
interest
Describe a data structure: Sequential Pattern
Tree
Mining algorithm to generate frequent episodes
Optimize SP Tree
Sequential Patterns
Pattern language
A = {A1,A2,….Am}
D1, D2,…,Dm = Domains
e over A is a (m + 2) tuple(a1, a2,..,am,
tbeg, tend)
Example of Events in the Stock Market
Domain
Event ID
Date
Comp Type
Comp Name Movement
e1
01/02/91
Computer
Microsoft
Down
e2
01/03/91
Computer
Microsoft
Up
e3
01/02/91
Computer
Microsoft
e4
01/03/91
Computer
Microsoft
Volatility
Activenes
Low
High
Medium
High
NoMovmt
High
Low
Down
High
High
Definitions
 Ordering Constraint
 Serial Occurrence
e -> f ,
e.tend < f.tbegin
 Parrallel Occurrence
(e || f)
 Attribute constraint
 Selection Constraint
e.type = ‘computer’
 Join Constraint
e.name = f.name
Event specification
Partial specifications
e[(e.type = ‘computer’ v e.type = ‘electronic’) ^
e.movement_direction = ‘down’]
comparing some characteristics
e[e.movement_direction = ‘up’] -> [e.name =
f.name] f[f. movement_direction = ‘down’]
Data Structure





Leaf node represents an event
An interior node represents an ordering constraint
If . is an ordering constraint labeling some interior
node, and if e and f are the left and right children of
that node then e . f is a sequential pattern.
Associated with each node is a table of matching
events
Attached to each node is a Boolean expression tree
representing attribute constraints
SP Tree
Matching
episodes
Matching
events
Matching
events
=
e
=
e.mvmt
e.name
up
f
f.name
f.mvmt
=
down
SP Tree for e[e.mvmt = ‘up’] -> [e.name = f.name]f[f.mvmt = ‘down’]
User specified pattern
Bottom-up algorithm
Intialize queue Q to empty
for (each leaf 1 in T) do begin
generate events from S that match constraints of 1
if(the parent p of 1 is not ready in Q) then
put p in Q
end
While (Q is not empty) do begin
Remove node n from Q
Generate_Events(n)
if(for n’s parent p another child was processed) then
put p in Q
end
Generate-events Algorithm
 for(each episode e from left child l of n) do begin
for (each episode f from right child r of n) do begin
if(node n is serial) then
if(e.tend >= f.tbegin) then
continue
if(events in e and f match the join constraint) then
form new episode g from events from e and f
end
end
Experimental evaluation
Results
window size variation
data set size
number of event specifications
attribute constraints
18
16
14
12
Time in Secs 10
8
6
4
2
0
4
5
6
7
8
9
Window Size in Days
Minimum Frequency = 0.8
10
11
40
35
30
25
Time in Secs 20
15
10
5
0
1
2
3
4
5
6
Number of Event specifications
Minimum Frequency = 0.8
Window size = 11
30
25
20
Time in Secs 15
10
5
0
3
4
5
6
Number of constraints
Minimum Frequency = 0.8
window size = 5
7
7
6
5
Time in Secs
4
3
2
1
0
2
3
4
5
6
7
8
9 10 11 12
Number of Events in Data sets
Minimum Frequency = 0.7
Window size = 5
SP Tree Optimization
If two event nodes represent the same
event, then only one of the nodes can be
used.
If two ordering nodes have the same join
constraints, and they both have the left
and right children representing the same
events then one such node is sufficient.
Conclusions
Approach is
Robust
Flexible
Efficient
Complex pattern
Good performance
References
Discovering frequent episodes in sequences by
Mannila. H., Toivonen, H and Verkamo
Agarwal, R., and Srikanth “Mining sequential
patterns”
Mannila. H., Toivonen, H “ Discovering
generalised episodes using minimal occurences
Agarwal, R., and Srikanth”Mining generalised
association rules