Transcript (AC) Mining for A Personnel Scheduling Problem

Associative Classification (AC) Mining
for A Personnel Scheduling Problem
Fadi Thabtah
Trainer scheduling problem
Schedule
Courses (events)
Locations
Staff (trainers)
Resources
Timeslots
Trainer scheduling problem
• Assigning a number of training courses
(events) to a limited number of training
staff, locations, and timeslots
• Each course has a numerical priority
value
• Each trainer is penalised depending on
the travel distance
Objective Function
Total priority for
scheduled events
Total penalty for training
staff
MAX
Hyperheuristic approach
• Operates at a higher level of abstraction
than metaheuristics
• You may think of it as a supervisor that
manages the choice of simple local
search neighbourhoods (low-level
heuristics) at any time
Low-level heuristics
•
•
•
•
Problem-oriented
Represent simple methods used by human experts
Easy to implement
Examples:
- Add new event to the schedule
- Swap two events in the schedule
- Replace one event in the schedule by another
Hyperheuristi
c
Current solution
Low Level
Heuristic 1
Low Level
Heuristic 2
Perturbed solution
Low Level
Heuristic 3
Building a Schedule using A
hyperheuristic
Initial
solution
Objective
value
Hyperheuristic
algorithm
CPU
time
Set of lowlevel
heuristics
Selected lowlevel heuristic
Objective
value
Perturbed
solution
Objective
value
Current
solution
(according to
acceptance
criterion)
Advantages of hyperheuristics
• Cheap and fast to implement
• Produce solutions of good quality
(comparable to those obtained by hardto-implement metaheuristic methods)
• Require limited domain-specific
knowledge
• Robustness: can be effectively applied
to a wide range of problems and
problem instances
Current Hyperheuristics
Approaches
• Simple hyperheuristics
(Cowling et al., 2001-2002)
• Choice-function-based
(Cowling et al., 2001 – 2002)
• Based on genetic algorithms
(Cowling et al., 2002; Han et al., 2002)
• Hybrid Hyperheuristics.
(Cowling, Chakhlevitch 2003-2004)
Why Data Mining
Scenario:
While constructing the solution of the scheduling
problem, the hyperheuristic manages the choice
of appropriate LLH in each choice point, therefore
an expert decision maker is needed
(Classification). Two approaches:
1. Learn the performance of LLH from past
schedules to predict appropriate LLH in current
one
2. While constructing schedule learn and predict
LLH Or what so called, Learn “On-the-fly”
Classification : A Two-Step Process
1. Classifier building: Describing a set of predetermined classes
2.
Classifier usage:
Classification
Algorithm
• Calculate error rate
• If Error rate is acceptable, then
apply the classifier to test data
Training Data
Test Data
RowId
A1
A2
1
2
3
x1
x2
x1
y1
y4
y1
Class
Classification
Rules
RowIds
A1
A2
Class/
LLH
1
2
3
4
5
6
7
8
9
10
x1
x1
x1
x1
x2
x2
x2
x1
x2
x3
y1
y2
y1
y2
y1
y1
y3
y3
y4
y1
c1
c2
c2
c1
c2
c1
c2
c1
c1
c1
Learning the Performance of LLH
Data Mining
Techniques
Applied K
times
Produce
(Hyperheuristic Solution)
llh
oldpriority
newpriority
oldpenalty
newpenalty
applied
1
71954
72054
793
790
1
2
71954
71954
793
793
0
20
71954
71054
793
761
0
27
71954
71954
793
793
0
37
71954
71954
793
793
0
43
71954
71954
793
793
0
47
71954
71954
793
793
0
58
71954
71954
793
793
0
68
71954
71954
793
793
0
74
71954
71954
793
793
0
Derived
Hyperheuristic
Algorithm
Guide
Rules Set
(If/Then)
Association Rules Mining
• Strong tool that aims to find
relationships between variables
in a database.
• Its applied widely especially in
market basket analysis in order
to infer items from the presence
of other items in the customer’s
shopping cart
• Example : if a customer buys
milk, what is the probability that
he/she buys cereal as well?
•
Unlike classification, the target class is
not pre-specified in association rule
mining.
Advantages:
• Items shelving
• Sales promotions
• Future planning
Transactional Database
Transaction Id
Items
Time
12
bread, milk, juice
10:12
13
bread, juice, milk
12:13
14
milk, beer, bread,
juice
bread, eggs, milk
13:22
beer, basket, bread,
juice
15:11
15
16
13:26
Associative Classification (AC)
• Special case of association rule that considers only
the class label as a consequent of a rule.
• Derive a set of class association rules from the
training data set which satisfy certain user-constraints,
i.e support and confidence thresholds.
• To discover the correlations between objects and
class labels.
Ex:
• CBA
• CPAR
• CMAR
AC Steps
Associative classification Algorithm
Frequent Ruleitems:
user
Attribute values that pass support
threshold
Class Association Rules
Training
Data
Rule support and confidence
Given a training data set T, for a rule R : P  c
• The support of R, denoted as sup(R) , is the number of
objects in T matching R condition and having a class label c
• The confidence of R , denoted as conf(R), is the the number
of objects matching R condition and having class label c
over the number of objects matching R condition
• Any Item has a support larger than the user minimum
support is called frequent itemset
Current Developed Techniques
•
•
MCAR (Thabtah et al., Pceeding of the 3rd IEEE International
Conference on Computer Systems and Applications (pp. 1-7)
MMAC (Thabtah, et al., Journal of Knowledge and Information System
(2006)00:1-21.
MCAR Characteristics:
MMACC characteristics:
 Combinations of two general
 Produces classifiers of the form:
data mining approaches, i.e.
v1  v2  ...  v  c1  c2  ...c
(association rule, classification)
that are suitable to not only
 Suitable for traditional
traditional binary classification
classification problems
problems but also useful to multi Employs a new method of
class labels problems such as
finding the rules
Medical Diagnoses and Text
Classification.
 Presents three Evaluation
Accuracy measures
k
i
Data and Experiments
Learning Approach : Learn the performance of LLH
from past schedules to predict appropriate LLH in
current one
Supp=5%, confidence=40%
Number of datasets : 12-16 UCI data and 9 solutions
Of the training scheduling problem
Algorithms used:
CBA (AC algorithm)
•
MMAC (AC algorithm)
•
Decision Tree algorithms (C4.5)
•
Covering algorithms (RIPPER)
•
Hybrid Classification algorithm (PART)
Relative prediction accuracy in term of PART
for the Accuracy Measures of MMAC algorithm
110.00%
90.00%
( AccMMAC  AccPART )
AccPART
70.00%
50.00%
30.00%
10.00%
-10.00%
1
2
3
4
5
6
7
8
9
-30.00%
-50.00%
Top-label
Label-w eight
Support-w eight
PART
Any-Label
Relative prediction accuracy in term of CBA for
the Accuracy Measures of MMAC algorithm
95.00%
Difference in Accuracy %
85.00%
75.00%
( AccMMAC  AccCBA )
AccCBA
65.00%
55.00%
45.00%
35.00%
25.00%
15.00%
5.00%
-5.00%
1
2
3
4
5
6
7
8
9
-15.00%
-25.00%
-35.00%
Nine Scheduling Runs
CB A
To p-label
A ll-label
A ny-label
Number of Rules of CBA, PART and Toplabel
Number of Rules
26
24
22
20
18
16
14
12
10
8
6
4
2
0
Run1 Run2
To p Label
P A RT
Run3
CB A
Run4
Run5
Run6
Run7
Run8
Ten Runs Scheduling Data
Run9
Co
Ti
c
nt
ac - Ta
t-l
c
en
se
s
Br
L
ea
st e d7
-c
an
W ce r
ea
th
He er
ar
t
He -c
ar
tLy s
M mp
pr ush h
im
r
a r oo m
ytu
m
or
Vo
te
CR
X
Ba
Si
la
nc
c
e- k
sc
al
e
Au
Br tos
Hy ea s
po t-w
th
yr
oi
d
z
kr oo
-v
skp
(%)
Accuracy (%) for PART, RIPPER, CBA
and MMAC on UCI data sets
100
90
80
70
60
50
40
30
20
10
0
PART
RIPPER
CBA
MMAC
Data sets
Comparison between AC algorithms on 12
UCI data sets
30
27
24
(%)
21
18
CBA
15
CMAR
CPAR
12
MCAR
9
6
3
0
1
2
3
4
5
6
Data
7
8
9
10
11
12
MCAR vs. CBA and C4.5 On UCI
data sets
Classifier Number of Rules
Dataset
MCAR
CBA
C4.5
MCAR
CBA
C4.5
Tic-tac
100
100
83.61
26
25
95
Balloon
100
100
100
3
3
3
Contact
100
93.33
83.33
9
6
4
Led7
72.32
71.1
73.34
192
50
37
Breast-cancer
71.61
69.66
75.52
71
45
4
Weather
100
100
50
6
6
5
MACR Support and Confidence Values for contact
dataset
1.2
1
Percentage %
Classification Accuracy %
0.8
Supp
0.6
Conf
0.4
0.2
0
1
2
3
4
5
6
7
8
9
R anked R ules
Heart-c
80.4
79.87
78.12
72
44
12
Heart-s
81.35
79.2
81.29
31
22
2
CBA Support and Confidence Values for contact
dataset
1.2
78.5
75.09
83.78
48
38
12
Mushroom
97.65
94.18
100
48
45
33
primary-tumour
40.5
25.47
42.47
28
1
23
Vote
90.1
86.91
88.27
84
40
4
CRX
83.05
85.31
80.72
97
43
54
Sick
93.88
93.9
93.87
17
10
1
Credit-Card
70.26
70.4
71.8
162
116
54
1
Percentage %
Lymph
0.8
0.6
Supp
Conf
0.4
0.2
0
1
2
3
4
Ranked Rules
5
6
Conclusions
• Associative classification is a promising approach in
data mining
• Since more than LLHs could improve the objective
function in the hyperheuristic, we need a multi-label
rules in the classifier
• Associative classifiers produce more accurate
classification models than traditional classification
algorithms such as decision trees and rule induction
approaches
• One challenge in associative classification is the
exponential growth of rules, therefore pruning
becomes essential
Future Work
• Constructing a hyperheuristic approach for the personnel scheduling
problem
• Investigating the use of multi-class labels classification algorithms
with a hyperheuristic
• Implementing of a new data mining techniques based on dynamic
learning suitable for scheduling and optimization problem.
• Investigate rule pruning in AC mining
Questions
?