Title: Multiobjective Clustering via Metaheuristic Optimization: An Application to Market Segmentation
1Multiobjective Clustering via Metaheuristic
Optimization An Application to Market
Segmentation
- Rafael Caballero
- Manuel Laguna
- Rafael Martí
- Julián Molina
2Clustering
i
aij
j
aij dissimilarity between i and j
3Dissimilarity Matrix
Attributes for object i
Attributes for object j
ri1, ri2, , rin
rj1, rj2, , rjn
4Multiobjective Problems
- Partitioning of objects using one partitioning
criterion but multiple dissimilarity matrices - Partitioning of objects using one dissimilarity
matrix but more than one partitioning criteria
5Objective Functions
- Partition diameter
- Unadjusted within-cluster dissimilarity
- Adjusted within-cluster dissimilarity
- Average within-cluster dissimilarity
6Illustrative Example
7Multiple Dissimilarity Matrices Market
Segmentation
- Data sources
- Performance drivers (descriptor variables)
- Performance measures (response variables)
- Find clusters that are homogenous with respect to
the performance drivers and that at the same time
help to explain the variation on the performance
measures
8Telecom Example (Brusco, Cradit and Stahl, 2002)
- Survey of 4400 business units
- Self-reported measures of firm technology and
telecom activity - LAN activity, desktop computing, remote access
services, network transport activity, premise
equipment inventory, telecommunication spending
estimates - Partitioning of market into homogenous clusters
of business that were interested in network
services and maintenance products
9Clustering Procedure
- Variable selection (from 21 to 9)
- Remote data communication points, LAN segments,
network size, trunks, toll-free-lines, high-speed
digital lines, desktop computers, LAN nodes and
network routers - Application of bi-criterion clustering simulating
annealing heuristic (SAH) - Identification of homogenous high-spending
segments
10Mathematical Model
for l 1, , L
Maximize
where
11Scalar Objective Function
Minimize
The principal limitation of multiobjective
programming is the selection of an appropriate
weighting scheme. The severity of this problem
increases markedly when three or more criteria
are considered. (Brusco and Stahl, 2005)
12Multiobjective Metaheuristic Approach
- Scatter/Tabu search hybrid
- Approximation of Pareto front in a single run
- Use of compromise programming principles to guide
the search - Density approximation (distance to RefSet) used
as stopping criteria
13Phase I Initial Tabu Searches
14Global Criterion Method (Compromise Programming)
Ideal Point
Approximated from
Anti-ideal Point
Use random weights until InitPhase searches fail
to produce a new efficient point
15Phase I Tabu Searches with a Global Criterion
2
x1
x2
5
x4
x6
1
7
6
3
4
x5
x3
16Phase II Scatter Search
- RefSet consists of
- Best single-objective solutions
- Diverse solutions (Max-min criterion using L?)
- An updated archive of efficient solutions is
maintained throughout the search
17Phase II Improvement Method
Efficient frontier
xi
Ideal (xi , xj)
f2
Compromise point for (xi, xj)
xj
New trial solution
Search area
f1
wi 1 for compromise point
18Phase II RefSet Update
- Choose best p solutions according to the
individual objective functions - For each solution x ? \TabuRefSet calculate a
normalized L? distance and find - Create a list of eligible points
- Choose b-p eligible points (sequentially) to
maximize
19Solution Representation
where xk (1 xk n) is the centroid of cluster
k
Then,
for k 1, , K
20Tabu Search Neighborhood
- Generate NSize (r, q) pairs such that q is not in
x and r is between 1 and K - Replace xr with q and evaluate the move with
either one of the objective functions (initial
p1 searches) or L? metric
21Combination Method
Repair trial solution when the same object
appears twice
22Branch and Bound Results
23Centroid-based Solution Representation
24Experiments with Multiple Criteria
25Experiments with Multiple Dissimilarity Matrices
26Application to Market Segmentation
27Conclusions
- Hybrid SS/TS approach seems to be an effective
approach to multiobjective optimization problems - Multiobjective problems in data analysis are a
fertile application ground for metaheuristic
optimization methods