Mining Customers

1
Mining Customers Preferences for Automotive
Design
Dr. Djamel Bouchaffra(http//www.oakland.edu/bou
chaff ) Intelligent Human Computer Interaction
LaboratoryComputer Science and Engineering
Department Oakland University
2
Overview

• Problem Description
• Perception Collection and Survey
• Classes Definition using WordNet
• Car External Design
• Car Contour Extraction-Training Phase
• Classifier Engines
• Car Internal Design
• Impact of Internal Design on Car Choice
• Internal Design Engineer Score

3
Problem Description
boxy
attractive
ordinary

4
Customers Profiles

• Vector profiles assigned to customers are built
  as follows
• If age ? 30 then age 1 else if (31 ? age ? 50
  then age 2 else age 3
• If male then gender 0 else gender 1
• If senior executive then SPC 0 else if middle
  manager then SPC 1 else if student then SPC
  2 else SPC 3
• If African then CB 1 else if European then CB
  2 else if North American then CB 3 else if
  South American then CB 4 else CB 4.

5

• The similarity S between 2 profile vectors C1
  1, 0, 2, 3t and C2 2, 0, 2, 3t is equal to
  ¾
• D(C1, C2) 1 S(C1, C2) ¼
• A distance of 0 means that the 2 customers are
  from the same profile

6
Similarity between Perceptions

• A perception is an adjective (unit of
  information-it can be extended to phrases)
• Customers profiles need to be incorporated in the
  perceptions similarity for accuracy
• Many techniques of distance computation between
  words exist, but they depend on external
  information
• The only information we have available are the
  customers opinions regarding an automobile

7
Similarity between Perceptions (cont.)

• Our idea is to incorporate the lexical database
  WordNet
• The similarity between a pair of customers
  perceptions that we have developed is
• Where Wk (p) W o W o o W(p) is a set of
  synonyms assigned to p by composing WordNet
  operator k times.

8
Similarity between Perceptions (cont.)

• Example Compute ?(attractive, elegant) from the
  following two customer profiles
• C11,1,2,4t and C23,0,0,3t
• D(C1, C2) 1 S(C1, C2) 1
• W(attractive) gorgeous, captivating,
  handsome, beautiful, exquisite
• W(elegant) handsome, neat, exquisite,
  refined, smart
• ?1 (attractive, elegant) ¼.This value is
  smaller than 0.37 which is the similarity between
  these two perceptions from the same customer
  profiles

9
Perception Collection and Survey
The side view of this car is attractive, the
front view is beautiful but the rear view is
ordinary.
Speech recognizer with embedded part-of-speech
tagger
beautiful attractive ordinary
beautiful
0.8
0.55
(similarity measures between adjectives)
attractive
0.50
ordinary
10
Class Definition via WordNet
captivating
mysterious
0.8
captivating
0.7
0.55
speedy
0.1
attractive
Perception graph extraction
performant
small
ugly
tight
narrow
boring
good
0.6
0.8
unconfortable
dull
elegant
gorgeous
0.7
performant
ugly
speedy
Kernels Extraction
Next Page
11
Class Definition via WordNet (cont.)
small
tight
captivating
mysterious
narrow
0.8
0.6
0.6
0.55
unconfortable
attractive
boring
0.75
dull
good
0.8
unconfortable
elegant
gorgeous
ugly
performant
0.7
speedy
?1 captivating ?2 narrow ?3 elegant ?4 boring
Concept Extraction
12
Class Definition via WordNet (cont.)

• An unsupervised clustering has been performed to
  extract the global feedback of customers
  regarding all automobiles
• From each kernel (cluster), we extracted the
  vertex ?? that is the closest to each other
  vertex
• The classes (or labels) for the classification
  task are these vertices ?s

13
Contour Extraction (using FD)-Training
Car Model (Chrysler Concorde)
3 Views
3 Contours
Attractive
Ordinary
Beautiful
Database
14
Classifier Engines
Car Model Chrysler Concorde (New input)
k-NN
SHMM
NN
Results
Ordinary
Attractive
Beautiful
15
Car Classification Program
16
Performances Table
Precision ()  Sample size kNN NN SHMM
70 cars 52.1 54.2 66.7
114 cars 73.2 78.6 89.3
SHMM is more adequate for customers view
prediction !
17
Classifier Engines Combination
Car Model (Chrysler Concorde)
k-NN
SHMM
NN
Results
Ordinary
Attractive
Beautiful
Multi-classifier Module
Attractive
Final Class
18
Open Research Issues

• Traditional classifiers tend to assign the same
  class to slightly different car shapes
• Does metrics fully account for categorization ?
• we believe the answer is no!
• Therefore, specific classifiers need to be
  developed in order to solve the slight
  deformation-switch problem

19
Internal Design

• Influence of options on customers decision
  cruise control, power windows, sound systems

20
Internal Design Engineer Score

• Option priority level from customers
• Strength of agreement of a priority (Borda Count)
• A penalty/reward scheme (customers drive the car
  design phase!).
• Option BC (from a survey)(cd, 0.8) (ls,
  0.75) (hs, 0.67) (cc, 0.55) (sr, 0.44) (pw,
  0.25)
• Engineers score
• 1 ? 0.8-1 ? 0.4-1 ? 0.551 ? 0.751 ? 0.671 ?
  0.25 1.48
• Max score 0.8 0.75 0.67 0.55 0.44 0.25
  3.46
• Final engineers score 1.48 / 3.46 42

21
Conclusion and Future Work

• Investigate other shape modeling techniques that
  better map perceptions/design (abrupt changes of
  contour)
• Develop an intelligent shape synthesis system to
  produce sequence of shapes through natural
  language commands using a virtual reality display
  
• Better understand categorization and metric
  relationship and their impact in automotive
  design through workshops that gather
  statisticians, psychologists and AI researchers

22
Some Related References

• 1. D. Bouchaffra and S. Abul-Hassan, Automotive
  Design Driven by Pattern Recognition, Smart
  EngineeringSystem Design, Annie'2001 Conference,
  University of Missouri-Rolla, November 4-7, 2001
  (Nominated for the best paper award at ANNIE2001
  Conference)
• 2. D. Bouchaffra and J. Tan, Mapping Designs to
  User Perception using a Structural HMM
  Application to Kansei Engineering, in
  International Conference on Computational
  Intelligence for Modeling Control and Automation,
  2003.
• 3. D. Bouchaffra and J. Tan, Structural HMM
  Modeling and its Applications in Automotive
  Industry, The 5th International Conference on
  Enterprise Information Systems, Angers France,
  23-26 April 2003.
• If you have any interest in this research, or
  need any material, please contact me at
• dbouchaffra_at_ieee.org