Computational Marketing Using An AppleSeed Cluster

1
Computational Marketing Using An AppleSeed Cluster
Academic Developers Conference 2001

• Dr Kevin Voges Dr Nigel Pope
• Griffith University

2
Overview

• Computational Marketing
• What is computational marketing?
• Application area - Rough Clustering
• Building a Computer Cluster
• Building an AppleSeed cluster
• How to use a computer cluster
• Programming considerations

3
What is Computational Science?

• Theoretical Science and Experimental Science
• Problems
• Difficult to use experimentation in the study of
  social and market processes
• Mathematics-based theories are not always the
  most useful method of presenting explanations of
  social and market behaviour
• Computational Science - third approach
• Benefits
• New forms of experimentation through study of
  artificial societies and markets
• New forms of theory development
• New forms of data analysis based on developments
  in Computational Intelligence.

4
Computational Science Literature

• Biology
• Environmental Studies
• Geography
• Organizational Theory
• Anthropology
• History
• Sociology and Social Science
• Political Science
• Economics and Finance.

5
What is Computational Marketing?

• The application of computational science
  techniques to modelling market behaviour and
  analysing marketing data

• can consider market mechanisms other than
  priceeg. the satisfaction of consumer needs
• can consider communication through social
  networks
• can explicitly study the role of memory in
  consumer behaviour
• offers a wider range of techniques for analysing
  marketing data.

6
Rough Clustering

• Based on Rough Sets theory (Pawlak)
• Rough Set defined by two sets
• Lower Approximation - object definitely in the
  set
• Negative region - objects definitely not in the
  set
• Boundary region - objects may or may not be in
  the set
• Upper Approximation - union of Lower
  Approximation and boundary region
• Rough Cluster defined in the same way
• Lower and Upper Approximation of cluster
• Membership in rough clusters determined by a
  distance measure.

7
Rough Sets
Person Age Grp Purchase Adam 16 -
25 Yes Bill 16 - 25 No Colin 26 -
35 No Dave 26 - 35 No Evan 36 -
45 Yes Fred 36 - 45 Yes George 36 - 45 Yes
8
Rough Set
Person Age Grp Purchase Adam 16 -
25 Yes Bill 16 - 25 No Colin 26 -
35 No Dave 26 - 35 No Evan 36 -
45 Yes Fred 36 - 45 Yes George 36 - 45 Yes
Selected subset of data Call subset X
9
Rough Set
Person Age Grp Purchase Adam 16 -
25 Yes Bill 16 - 25 No Colin 26 -
35 No Dave 26 - 35 No Evan 36 -
45 Yes Fred 36 - 45 Yes George 36 - 45 Yes
Lower approximation of X based on Age Group
10
Rough Clustering
Person Age Grp Purchase Adam 16 -
25 Yes Bill 16 - 25 No Colin 26 -
35 No Dave 26 - 35 No Evan 36 -
45 Yes Fred 36 - 45 Yes George 36 - 45 Yes
Upper approximation of X based on Age Group
11
Rough Clustering

• Computationally expensive
• Large data set (6000 records)

• Number of potential solutions
• Need to form clusters which maximise size of
  lower approximation and minimise size of boundary
  region
• Solution - Evolutionary Algorithm using a
  parallel computer cluster

12
Rough Clustering

• Evolutionary Algorithm

• Uses simple model of evolution to find solution
  to problems
• Population (ie. more than one) of solutions
  randomly generated
• Effectiveness (fitness) of each solution
  calculated
• New solutions are produced by combining and
  randomly changing older solutions (mutation)
• Best solutions kept each generation
• Process continues until some specified stopping
  criteria reached.

13
Building a Computer Cluster

• Cluster
• group of computers connected by a switch
• Beowulf Cluster - Unix system

• Building an AppleSeed Cluster
• Obtain hardware
• Connect hardware
• Install software
• Configure software

14
Building an AppleSeed Cluster

• Obtain Hardware

• A number of Macintosh computers with built-in
  Fast Ethernet
• A Fast Ethernet switch containing a port for each
  Macintosh
• A corresponding number of Category 5 Ethernet
  cables with RJ-45 jacks

15
Building an AppleSeed Cluster

• Connect Hardware

• Connect one end of each Ethernet cable into the
  Ethernet port on the Macintosh
• Connect the other end into the port on the switch
• Connect monitor to at least one computer
• Connect power cables, etc

16
Building an AppleSeed Cluster

• Install Software

• Autoguest control panel on each computer
• AppleSeed Folder on all computers startup disk
• Launch Puppy on each computer
• Launch Den Mother on main computer

17
Building an AppleSeed Cluster

• Configure Software

• Mac TCP/IP set to Ethernet (Using DHCP Server)
• Turn File Sharing and Program Linking on
• File Sharing Users and Groups - Allow guests to
  connect - Allow guests to link to programs
• Energy Saver to Never Sleep

18
Building an AppleSeed Cluster

• Current Configuration

• Five G4s
• Cisco Catalyst 3600XL 12-port Ethernet Switch
• MasterView Monitor Switch - can view and control
  the desktop of four G4s from a single
  monitor and keyboard

19
Building an AppleSeed Cluster
20
EP2 Cluster

• Fastest AppleSeed cluster in the world?
• 16 Dual PowerPC G4/450
• 50 GigaFlops
• Located in Portugal
• Used for numerical simulation of plasmas

21
How to Use a Computer Cluster

• Run same program with different parameters on
  each computer
• Set up a master / slave process between main
  computer and other computers
• Divide large matrix into subsections and have
  each computer process a subsection

22
Same program with different parameters

• Market models using simulation require
  investigation of the effect of different
  parameters(eg. size of individuals social
  network, speed and depth of word-of-mouth
  endorsements)
• Computer cluster allows for same programs to run
  independently on each computer with changed
  parameters
• Multiple computers enables faster exploration of
  effects of varying values of parameters.

23
Master / slave process

• Main activity in evolutionary algorithm is
  assessing fitness of individual solutions
• Depending on type of problem, this can require
  noticeable real time
• Solution- master computer controls process and
  sends copy of solution to slave computers-
  slave computer calculates fitness for solution
  and returns fitness value to master computer
• Process continues until stopping point reached.

24
Master / slave process
25
Divide matrix into subsections

• Some problems require manipulation of large
  matrix of values
• Solution- divide matrix up into sections- send
  section to each computer- computer processes
  section and returns result to main computer-
  main computer recombines sections into full
  matrix and shows solution
• More common in physical sciences.

26
Programming Considerations

• Use MPI (Message Passing Interface) to control
  computers
• MPI is library, not a language
• The computers coordinate their activities by
  explicitly sending and receiving messages
• MPI is large and small
• standard implementation has 125 functions
• can write a useable program with 6 functions.

27
Programming Considerations

• Minimum MPI
• Initialise MPI
• Find out how many computers there are
• Find out which computer I am
• Send a message
• Receive a message
• Terminate MPI
• MacMPI
• public domain version from UCLA
• subset of full MPI.

28
Conclusion
In the applications described in this paper, the
choice of computer platform has enabled
researchers, with an extensive knowledge of
market structure and consumer behaviour, but with
limited knowledge of parallel computers and
computer networking, to access very powerful
computer technology.
29
QA
30
AppleSeed Project Web Site
http//exodus.physics.ucla.edu/appleseed/appleseed
.html
31
Computational Marketing Using An AppleSeed Cluster
Academic Developers Conference 2001

• Dr Kevin Voges Dr Nigel Pope
• Griffith University