Automated Conceptual Abstraction of Large Diagrams By Daniel Levy and Christina Christodoulakis December 2012 (2 days before the end of the world)

 Introduction  Big picture  Clustering Algorithm  Experiment & Results  Conclusion Outline

 Introduction  Big picture  Clustering Algorithm  Experiments & Results  Conclusion Outline

 So what is this “clustering” you speak of?  Why do we need to cluster?  Reduce cognitive load Introduction

 Introduction  Big picture  Clustering Algorithm  Experiment + Results  Conclusion Outline

Big Picture

Vision

Diagram Abstraction

 Its been done before.. Related Works

 Consider a diagram stripped of semantics, or pre processed using methodologies in previous work  Cluster graph  Evaluate clusters proposed based on closeness of meaning in the node names Our Approach

 Introduction  Big picture  Clustering Algorithm  Experiment + Results  Conclusion Outline

Min-Cut

Naïve Min-Cut Algorithm

C A N B C A N B 2 3 E 4 E 4 *Must result in exactly 2 partitions Combinations / Creating partitions *Assume there exist additional nodes

C A N B C A N B 1 E E 4 4

C D A B C D A B 2 Minimum sets C D A B C D A B 2 3

D A B D A B 3 2 D A B D A B 2 Cycles

E D C A B Listing the min-cuts

E D C A B

E D C A B

E D C A B

E D C A B

E D C A B E D C A B Outside-in approach

E D C A B E D C A B

E D C A B E D C A B

E D C A B

 We use RiTa WordNet getDistance() function  We calculate pairwise distances between nodes.  Select for each node the smallest distance between it and another node  Sum all minimum distances  Average over all nodes in candidate cluster Cluster Distance Measure

 Introduction  Big picture  Clustering Algorithm  Experiments + Results  Conclusion Outline

Experiment 1

Experiment #1

User 1 abstraction

Experimentation User 2 abstraction

Experiment # 1 automated abstraction

Experiment 2

Experiment #2

Simplified version

 Introduction  Big picture  Clustering Algorithm  Experiments + Results  Conclusion Outline

 Surprised at how similar manual clustering and automated clustering were.  Suggested improvements:  Automatic distance threshold  Creating subgraphs  Strictness of clustering (min # of clusters  Advanced min-cut discovery Conclusions

Questions? Merry Christmas!