Visual Clustering with Artificial Ants Colonies

1
Visual Clustering with Artificial Ants Colonies

• N. Labroche, N. Monmarché and G.
  VenturiniLaboratoire d'Informatique de
  l'Université de ToursÉcole Polytechnique de
  l'Université de Tours Département
  Informatique64, avenue Jean Portalis 37200
  Tours, Francelabroche,monmarche,venturini_at_univ-
  tours.fr

2
Talk overview

• Goal
• ant-based clustering algorithms
• the chemical recognition system of ants
• Main principles of our model
• Visual AntClust algorithms
• Results and example
• Conclusion

3
Goal

• Building a visual clustering tool
• Idea
• Real ants solve a clustering problem in their
  everyday life ? nestmates recognition mechanism
• Method
• Modelling the chemical recognition system of real
  ants
• Extracting its main principles to create a new
  unsupervised clustering algorithm

4
Clustering Problem
5
Ant-based clustering algorithms (1/3)

• Brood sorting Lumer and Faieta (1994)
• Discrete grid on which ants move, pick up or drop
  randomly placed objects
• Problem two contiguous sets of objects can be
  considered as only one set
• AntClass Monmarché (2000)
• Hybridisation with k-Means
• Several objects on the same place

6
Ant-based clustering algorithms (2/3)
7
ant-based clustering algorithms (3/3)

• Topic maps for Web pages J. Handl (2002)
• Behavioural switches ("Eager ants", "Jumps")
• Acluster V. Ramos (2002)
• objects in the neighborhood ? Pheromones
  trails

8
Main principles of the chemical recognition
system of ants
Cuticular odour or  label  (hydrocarbons)
Neuronal template
Recognition phenotype matching mechanism ?
comparison between label and template
A set of behavioural rules (aggression, reject,
feeding, social licking, trophallaxy, )
Genome
9
Model of the chemical recognition system of ants
Satisfaction estimator s
Label 2D-vector
Template Acceptance threshold
Acceptance mechanism
Behavioural rules "Meeting" algorithm
Genome one object of the data set
10
Template learning principles

• Each ant a performs NL meetings
• Mean similarity
• Maximal similarity
• Template for ant a is defined as

11
Acceptance mechanism

• Acceptance between 2 ants a and b

12
Visual AntClust Main Algorithm
Initialize N ants
While NbIter iterations are not reached
Draw N ants in the 2D-odour space
Meeting(a,b), a,b randomly chosen ants
Repeat N times
End While
Group in the same nest all the ants within a
perimeter of value Dmax
Delete the nests that are too small
Reassign the ants with no more nest
13
Meeting (Ant a, Ant b)
D ? Euclidian distance between Labela and Labelb
D lt (1-max(sa,sb)) And Acceptance(a,b)
Yes
No
Increase sa and sb ? ants a and b are well-placed
Update Labela(b) according to Ra(b)
End
14
Experiments

• Visual AntClust is compared to
• K-Means
• AntClass
• AntClust an other ant-based clustering algorithm
  inspired by a discret modelling of the chemical
  recognition system
• 50 runs for each method and each data set

15
Data sets
Name Objects Attributes Clusters
Art1 400 2 4
Art2 1000 2 2
Art3 1100 2 4
Art4 200 2 2
Art5 900 2 9
Art6 400 8 4
Iris 150 4 3
Glass 214 9 7
Pima 798 8 2
Soybean 47 35 4
Thyroid 215 5 3
16
Clustering Error Measure
c expected cluster label c computed cluster
label
17
Results (1/2)
18
Results (2/2)
19
Example 1
Step 1
20
Example 1
Step 2
21
Example 1
Step 3
22
Example 1
Step 4
23
Example 1
Art 6 data set
24
Conclusion

• Visual AntClust is able to treat from little to
  important data sets
• It performs well and even better than k-Means
  initialised with the expected number of clusters
  for some data sets
• Perspectives finding automatically the best
  parameters setting
• www.antsearch.univ-tours.fr

25
The End
26
Other ant-based clustering algorithms (2/3)










1st group objects
2nd group objects
Artificial ants
Rs